Learning in Vision

Mysore park workshop on computer vision

2011-12-22T11:41:00.004-05:00

I spent the last two days in Mysore for a workshop on computer vision. It was great interacting with the leaders in computer vision -- AZ, Jitendra, Martial, Shmuel, Michal, etc. -- in a small group. Prior to this, I only had brief interactions with some of them at conferences. The talks were OK, with most of the people presenting a recent conference talk or a submitted paper. I liked the presentations by AZ, Pushmeet, Tamara, Alex, and Dhruv. Although I found the discussions around the presentations a bit lacking, but the conversations over coffee, lunch, and dinner were very enjoyable. Another notable event was the banquet party: food, booze-n-brains, and dancing -- who can complain?

There was a slight panic moment when I realized I was losing my voice and were to give my talk first thing next morning. Perhaps, I was having too much fun at the banquet -- we won the "happy table" prize :D. Luckily my throat was back to normal by the morning. (Moral: Never ever compromise on the fun in a party).

I prepared a presentation to discuss the open challenges in face detection (slightly modified slides with some non-public content removed can be found here). Please let me know if you have any feedback on this presentation. Note that some of the background work was borrowed from Cha Zhang's recent survey on face detection.

Google scholar citations

2011-11-20T07:18:00.001-05:00

Google scholar citations is open to public. Finally, we don't really have to make our own publications page -- DBLP had limited coverage over publication sources, and Microsoft academic search never really moved beyond the "proof of concept." It is an easy way to track citations, and the authors can "manually" merge multiple revisions of their papers -- its great!

This makes me wonder why Microsoft did not improve academic search, at all. I have a feeling that it would be really difficult to convince product managers to invest resources in something that concerns only the scientific community (which is tiny compared to the general Internet users). But then, why did researchers even bother making the prototype go out in the wild? -- just publish the related technology as a paper and leave it there. And if they did come this far, would it not make sense to persist. Imagine all the annotated data for matching citations one can get from such tools. All of us check the state of citations of our papers and would fix any errors when it comes to the references of our own work. Well, Google will soon have lots of annotated data on citations and references.

Do we still need citation resolution techniques?
Does this mean machine learning lost more ground to information retrieval?

NIPS reviewing

2011-08-07T02:54:00.000-04:00

The papers assigned to me for reviewing for NIPS this year were no different from the previous years. Each of these papers was a mixed bag of good ideas, decent experiments, but unjustified conclusions and tall claims. I am finally trying to formulate a personal wishlist for things to see and evaluate in a paper so I am more consistent in my reviews. Here are some of these:

Overall:

Identify the novelty or USP and be explicit about it. I want you to tell me where we are going to go with this, as opposed to have me figure out on my own. Trust me most of the reviewers are not looking for papers reviews as sources of homework exercises.
Please do not overstate your contribution. I can point out a few researchers that always do this, but that would be a career suicide :D. Instead of overstating your contributions, I prefer seeing a honest definition of the scope or domain where the proposed solution works.

Introduction:
Give the problem definition, a motivating scenario, and a high-level description of the proposed solution.
In one of the papers I reviewed this year, the solution was first introduced in Section 3. The first two sections were just hand-waving. I am pretty sure most of us would want this to be fixed in the future revisions of this paper. Even if the results were great, I am setting my reviewing standards to make sure they are presented well.

Theory:
Put the derivations in the appendix.
A paper is not a textbook chapter. I would rather get the whole idea before I delve into the necessary derivations.

Experiments:

Do proper comparisons.

Please do not include a benchmark from a decade ago, particularly if the state-of-the-art for that problem is much better now. This is one of the major issues with computer vision papers at NIPS -- people are still comparing face recognition results to Eigenfaces, and reporting face detection results on MIT-CMU data sets.

Cite related work with proper description of connections and distinctions.

Tell me how does a citation in the form of [12,17, 28] help the reader. We can all search related popular work using a search engine. I will appreciate if you give me a sentence describing the connection between your work and [12], [17], and [28]. Otherwise, I am going to assume this is either a self-citation, a suck-up attempt to potential reviewers, or you have actually not read this paper.
Even if you do not include experimental comparisons to some of the related work, please describe why you did not. Often there are good reasons not to include those -- e.g., too complex to implement just for the comparison's sake, not enough detail to implement, and differences in the problem setup. But please tell those to readers. If you do not, then I am just going to assume you are trying to cheat.

Provide relevant implementation details.

Everyone says this, but many of us ignore it in our own papers -- sometimes unintentionally. I feel that when we are writing a paper, we are so familiar with our work that we often (unintentionally) skip some minor details that may be essential to reproduce our research. Although it is, in part, the reviewers' responsibility to point this out, it is seldom the case in the current reviewing model for conferences with little/no provision for revisions. My suggestion is to have an external reader look at the experiments section before submitting the paper. The good news is if I am the reviewer (and I believe there are plenty others) you will get brownie points that add up to the final decision.

This time I used a point scheme based on these (+1/-1) features to decide on the overall recommendation for the papers I reviewed. I wish the conferences had the above points as explicit questions in actual review as it would probably serve two purposes:

The reviews are more objective and comprehensive at least on these aspects of the submission.
The authors get a direct feedback on which parts they did a terrible job and should improve upon.

Notes on graph cuts

2011-06-08T13:50:00.000-04:00

Finding the min-cut in a general graph is easy (see Ford-Fulkerson and related algorithms). The resulting min-cut may often lead to trivial cuts that puts a single node in one of the two partitions. In these cases, we alter the optimization by including terms for "balancing" the size of the resulting partitions (e.g., ratio-cut and N-cut). This problem of "balanced" min-cut turns out be NP-hard (for the known formulations of balancing).

A common way to approach an approximate solution for this problem uses two steps: (a) relax the optimization by allowing non-discrete cluster assignments (often solvable as normalized/un-normalized spectral clustering); (b) cluster these continuous assignments to obtain the discrete assignments.

Note that here are no guarantees about the quality of the solution of the relaxed problem compared to the exact solution (e.g., the case of the cockroach graph). See Spielman and Teng (1996) and Kannan, Vempala, and Vetta (2004) for more analysis.

Ulrike von Luxburg wrote an excellent tutorial on spectral clustering, which explains most of the above points very clearly.

Latex tip: fixing incorrect figure numbers in \ref{}

2011-04-11T12:33:00.002-04:00

Weird, but true.

Always use \caption{} before \label{} in \figure{}. Apparently changing the order can lead to incorrect figure numbers while using \ref{}

WWW'11 at Hyderabad, India

2011-04-04T00:48:00.003-04:00

This was my first time at WWW. (I uploaded my presentation slides on http://vis-www.cs.umass.edu/~vidit/presentations.html)

My take-home message from this conference was: WWW is all about solving a real problem. Unlike papers at CVPR, NIPS, etc. WWW papers are about clearly defining a (sometimes new) problem, convincing the readers that it is worth solving, and then solving it at a large scale. In other words, the readers want to be convinced that the solution works. The emphasis is on the importance of the problem and the effectiveness of the solution. Nobody cares if you used twenty quadratic and five variational terms in your formulation to achieve a particular solution. If your solution doesn't outperform simple logistic regression, its no good. Also you would be asking for trouble if you don't have error-bars, t-test, and/or an analysis of your results to explain where/why it works and where/why it doesn't.

That's why I liked WWW -- its about simple/elegant innovation, clever engineering, and thorough experimentation. One aspect that I found missing was theory (duh! this is a conference about experiments) Btw, papers related to advertising and auctions did have theory component, but I wasn't interested in them.

Some papers I found interesting:
1. Like like alike -- Joint friendship and interest propagation in social networks
2. Efficient K-Nearest Neighbor Graph Construction for Generic Similarity Measures
3. Supporting Synchronous Social Q&A Throughout the Question Lifecycle
4. Learning to Model Relatedness for News Recommendation
5. The Web of Topics: Discovering the Topology of Topic Evolution in a Corpus
6. A Word at a Time: Computing Word Relatedness using Temporal Semantic Analysis

Overall, I had a great time in Hyderabad.

Self promotion: two new papers

2011-02-15T13:04:00.003-05:00

The year 2011 has been great for me so far! Two of my papers that I like the most (so far) have been selected for oral presentations at the most suitable venues (in my opinion).

My paper on image re-ranking (Learning to Re-Rank: Query-Dependent Image Re-Ranking Using Click Data) will be presented at the WWW conference in March. This paper proposes a super-simple approach for re-ranking for tail (infrequent) queries for which the click data is sparse and unreliable. We show big improvements in performance over Bing image search. (This work was done during an internship at Microsoft Research.)

The other paper (for which my excitement is even more) is on face detection (Online Domain-Adaptation of a Pre-Trained Cascade of Classifiers). It will be presented at CVPR in June. The idea is simple -- all of the faces appearing in a single image share the scene characteristics, so the easy-to-detect faces can help analyze the difficult-to-detect faces in an image. To implement this idea, we present an algorithm for adapting the Viola-Jones classifier to a single image. Using the proposed algorithm, we obtained a huge improvement in detection performance on the FDDB data set.

Embedding "all" fonts in a pdf on Max OS X

2011-02-01T15:21:00.003-05:00

There are always some vexing issues with the IEEE/ACM compatibility for the camera-ready pdf of paper -- the most common being non-embedded fonts. In past I had access to a Windows machine with Adobe Distiller on it, which I used to fix this issue. Now I only have a Mac.

I tried the unix command line ps2pdf on a reasonable postscript generated using:
"dvips -P pdf -t letter -G0 -o A.ps A.dvi"

"pdffonts A.pdf" shows lots of non-embedded fonts (TrueType: Helvetica, Times, etc.)

The solution that worked for me was:
Open A.ps in Preview (it will try to convert) and then save as A.pdf

I verified that all of the used fonts were embedded using pdffonts (alternatively in File->Properties in Adobe Reader).

Btw, pdffonts is not on macports. I found it as xpdf-pdffonts after installing xpdf.

PS. I was working on the WWW camera-ready paper on image re-ranking :)

ICVGIP'10

2011-01-15T12:19:00.002-05:00

Posting a bit late :(

Last month I attended the Indian conference on computer vision, graphics, and image processing (or ICVGIP) in Chennai. The best thing I liked about the conference is that the young attendees (students) were full of enthusiasm. The not-so-positive thing I noticed is a seemingly little interest from several of their mentors.

I enjoyed all of the invited talks, particularly those by Alyosha Efros (CMU) and Aaron Hertzman (Toronto). There were a couple of interesting talks from IIIT-Hyderabad (students of PJN and Jawahar) and another by Pushmeet Kohli (MSR-Cambridge) on interactive segmentation.

A big difference I noticed between the nature of presentations I saw at ICVGIP compared to elsewhere (also confirmed by a few well-established international researchers) was: the talks had too much stuff (read unnecessary details) in it. It felt like the presenter did not care if the audience understood the idea/experiments but the objective was to show off a particular result or the breadth of her knowledge or the obfuscated mathematical expressions. (To avoid the risk of pissing off some individuals, I would refrain from citing specific instances). Later while discussing this issue with a senior researcher, we sadly acknowledged this issue as a product of our Indian upbringing with a focus on "proving ourselves." However, we were hopeful that things are changing and perhaps with better exposure the young researchers will focus on "educate others" in their presentations.

Gaussian process regression code

2010-11-17T10:06:00.003-05:00

I graduated a couple of months ago. After passing time changing the city, state, country, and continent and meeting a few deadlines, I have decided to clean-up some of the code for my thesis work and extract useful (I think) modules to share.

As the first installment, I have put up the C++ (modular) version of GP regression that uses lapack routines for the matrix operations. You need to use clapack distribution to use this code. The code can be found at http://vis-www.cs.umass.edu/~vidit/Code/GPR.tgz
Note that this code is inspired by Rasmussen's distribution.

There will be more code coming soon (well depends on my job).

Please let me know if you face any issues using this code.

ECCV acceptance rates (contd.)

2010-06-13T15:09:00.002-04:00

Thanks for the corrections in my previous post, Timo.

Aah.. that is what happens when you write a blog-post in three separate sittings :( Nevertheless, the point about the range of acceptance rates is still evident here -- e.g., Stereo ( 7.9% ) and Shape representation and recognition ( 37.5% ).

Related over-all statistics for area-wise acceptance rates:
mean a : 26.69
median : 25.7
std. dev.: 8.09

Categorizing the research areas using their acceptance rates:

0-20 %: Stereo; medical image analysis.

20-25 %: Segmentation and grouping; video and event categorization; motion and tracking; matching, registration, alignment.

25-30 %: Face, gesture, biometrics; image features; statistical models and visual learning.

30-35 %: Computational imaging; object and scene recognition.

>35 %: reflectance, illumination, color; multi-view geometry;
shape representation and recognition.

I wonder how many of the submitted papers were rejected because of "of limited interest/out of scope" ratings. I have a feeling that there would be some "good" papers in the < 20% category that were rejected because of this reason. And it is not the fact that the stereo papers were of worse average quality than other papers.

Would it not be great if all the reviews (or at least area chair's consolidation reports were made public (anonymously, perhaps). That would serve two purposes:
(1) more transparency and accountability to the reviewing process, and (2) identification and possible redefinition of the scope of the computer vision conferences.

Acceptance bias at computer vision conferences

2010-06-11T11:46:00.001-04:00

The statistics for the number of submissions and acceptances have been published on this year's ECCV website. They have broken down these statistics according to the research areas, which is good and highly informative. One thing that stood out to me, and which I found very surprising, is the relative difference between the acceptance rates for "object and scene recognition" and any other research area. To me, there are two possible explanations for this fact:

1. The average quality of papers submitted in object recognition is very different from the quality of papers submitted in other categories. IMHO, such a claim is ridiculous -- and I am pretty sure there are a LOT of people who would empathize with this claim. And I say that based on my tiny experience looking at the accepted papers at the CVPR, ICCV, and ECCV from the last few years.

2. The explanation that I find more plausible, and also alarming, is the obsession of computer vision community with object recognition. I would argue that, barring a few exceptions, most of the published work in object recognition at these computer vision venues is an application of existing statistical and machine learning techniques -- with neither any real insights being developed, nor any improvements in the state-of-the-art for a particular application. Note that, by an application, I mean a generalizable solution, and not the state-of-the-art performance on a particular data set.

Another thing to notice is that there is no category for "Applications" at ECCV this year. Where should the papers about image-search (on the internet) and cool cameras go? Should papers on human vision go under "Statistical models and visual learning?"

After looking at these statistics, I was wondering about the impact of this bias towards specific areas of research. Although it is commonly said and heard that when it comes to job search (fresh out of PhD), the quality of the papers matter much more than the quantity. I would argue that anyone who has been on the job market would agree that this statement is untrue in almost all situations. For considering a candidate, research labs typically have a criterion based on a minimum number of publications in top conferences, and I better not comment on academia. Looking at the ECCV acceptance rates, it appears that there is a huge difference in the standards for having a paper accepted at a computer vision conference across the sub-fields of computer vision. For instance, both object recognition and segmentation are fundamental and "popular" areas of research, but they have 20.5% and 8.7% acceptance rates, respectively. For people working on less "hip" areas for a PhD, this bias could significantly hurt their prospects. This effect is probably more severe for people who are looking to make a transition to areas outside computer vision.

I think this is a great initiative that ECCV organizers took to make such detailed statistics available. Perhaps, we need to re-organize computer vision conferences -- perhaps split into smaller conferences focused on specific areas. Currently, CVPR, ICCV, and ECCV get way more attention than focused conferences like ACM MM. Acknowledging the large scope of problems and approaches in computer vision, we as a community need to change our perspective. So next time I am looking at someone's list of publications, I will pay a little more attention to the kind of problems that person is working on as well. Hopefully, this will catch on!

ECCV Workshop on face detection

2010-04-06T00:01:00.003-04:00

I am organizing a workshop titled "Face Detection: Where we are, and what next?" at ECCV this year.

We have so far been able to get hold of Yann LeCun (NYU) and Hartmut Neven as invited speakers. Also, Microsoft Research India has kindly agreed to sponsor the best public implementation and best paper awards for our workshop.

Please spread the word.

Assignment problem: Hungarian algorithm

2010-02-20T16:03:00.004-05:00

Suppose there are n persons and m jobs. Also, for each person-job pair (i<=n, j<=m), there is a predefined benefit C_ij associated with it. The assignment problem refers to the determination of the assignment of persons to jobs such that: (a) the benefit is maximum, (b) every person is assigned to at most one job, and (c) every job has at most one person assigned to it.

I encountered this problem while evaluating the performance of face detection algorithms. The output of a face detection system is a set of image regions hypothesized as face regions. When multiple detections are hypothesized for a given image, the set of detections needs to be matched against the set of annotated image regions. This setting is similar to the above-mentioned assignment problem, where the benefit matrix is computed using some matching score for each detection-annotation pair. A commonly used score is defined as the ratio of set-intersection and set-union of the two image regions in the given pair.

I expect this problem to be manifested in a wide range of application domains, and hence expected to run into some standard algorithms and their implementations. I did find an algorithm -- Kuhn's Hungarian algorithm (and later improvements by Munkres et al.). Considering the broad applicability of this algorithm, I was a little disappointed with the unavailability of useful implementations, though (I think it would be useful to have this algorithm be included in the Algorithms books/courses). I found one implementation developed by Brian Gerkey. It has some minor bugs. After some performance improvements and testing, I will soon upload the modified code.

If anybody knows of a much faster solution to this problem, please let me know.

My past presentations

2010-01-02T17:26:00.002-05:00

I uploaded all the talks/presentations that I could excavate from my hard drive at
http://vis-www.cs.umass.edu/~vidit/presentations.html

Hopefully, someone will find them useful.

Practical AI

2009-12-04T23:50:00.004-05:00

On December 1st, I gave a guest lecture in Professor Allen Hanson's Artificial Intelligence class. Originally, I thought about presenting something related to my research, computer vision, or machine learning. After looking at what topics they covered in the class, I decided against all of these options and chose something completely new to me -- classical AI. I decided to talk about some practical, non-book-ish applications of search and planning. The first two things that came to my mind were: maps and travel planning. Well, the temporal vicinity of Thanksgiving, NIPS travel, and holiday season must have something to do with them too :D.

When I first started making an outline for my talk, I started getting a little nervous because I wasn't sure of: (a) how much of AI -- and not Algorithms -- would go in these applications, and (b) if I would be able to convey anything useful without talking about learning-based approaches.

I came across an article Computational Complexity of Air Travel Planning, by Carl de Marken. This article presents an excellent analysis of the issues that makes the problem of determining the cheapest flight really hard (undecidable). Reading this article not only made me appreciate this problem and related solutions, but also gave me a strong argument against the people who question the utility of studying (computability/complexity) theory for AI graduates. I found this result pretty cool and since then, I have been telling other people about this result and its implications.

Anyways, I ended up making a theme of "Holiday travel" for my talk with the following steps:
* How/when to travel? (Travel planning)
* Get around? (Maps)
* What to do? (Photosynth/CitySense)

My goal was not to present technical details, but to motivate the undergraduates to appreciate the "actual" AI applications that they encounter in day-to-day life -- not the 'AI in games' kind of artificial intelligence.

I would be very happy to receive any feedback or comments about my slides.

Writing papers

2009-11-20T22:20:00.003-05:00

I was writing a paper with another amazing co-author. At one point, a discussion about reviewing came up. Agreeing that there is a good chance that one of the reviewer may not be able to put in a reasonable amount of effort while reviewing the paper. So essentially there were two options: (a) focus on pleasing this third (at best)lazy reviewer, or (b) focus on the (hopefully) other two reviewers. For the first time, I realized that these two goals can be competing with each other. In particular, a careful reviewer would appreciate (in fact, may demand) certain details about an approach, whereas these details may easily annoy an inept reviewer. Although one can argue that the main aim of writing a paper is not to get through the review process, but to spread the research ideas, but it is a fact that to spread these ideas, you have to go through this review process. Do good work, but at the same time, understand the process of putting the work out there, and be willing to spend some effort to presenting it in effective way.

Leaving the general ethical preaching aside, we did end up taking the path of focusing on the "good" reviewer(s), and ignore the "bad" reviewers. I sincerely hope we don't end up getting disappointed.

One more thing that I have been hearing very often these days: "All the reviews are being done by immature graduate students, who are looking for reasons to reject papers." A call to all the grad students, who might end up reading this -- Folks, its up to us that we preserve our respect. do a good job reviewing yourself, and encourage fellow grad students to do the same.

Hiatus

2009-07-23T13:34:00.002-04:00

Somebody mentioned to me that its been quite a while since I posted anything on my blog. I have been a little busy with wrapping up two projects during the Spring semester, and now, I am spending my summer at MSR India. There have quite a few things I thought I would post "once I get home," but that never happened.

CVPR and ICCV this year should have something to look at, right?

Compact representation for "large-scale" computations

2009-04-19T16:00:00.003-04:00

Last Friday, I gave a presentation on some of the really interesting and recent work by John Langford et al. in the advanced machine learning seminar at UMass Amherst. After listening to John at a recent Machine learning friends lunch seminar, I got the high-level idea of this work so I went on to explore its details. Although this work addresses issues tangential to the focus of the seminar, it worked out fine, I guess.

I posted my slides at http://vis-www.cs.umass.edu/~vidit/presentations/repLargeScaleComp.pdf.

Please let me know if I misinterpreted/missed something important.

Infinitely Imbalanced Logistic Regression

2009-03-22T11:38:00.002-04:00

Art Owen had an interesting, if not surprising, paper titled Infinitely Imbalanced Logistic Regression in JMLR 8 (2007). In this paper, Owen investigates logistic regression for binary classification tasks where one of the classes have finite number of samples and the number of samples for the other class approaches infinity, thus creating an infinite imbalance between the two. He shows that although the intercept of the learned logistic regression approaches -infinity, the remaining coefficient vector approaches a non-trivial and useful limit. Furthermore, the minority class affects the coefficient vector ONLY through its empirical mean, thus it suffices (within sampling uncertainty, that is) to replace all the samples of the minority class with a single sample for the mean.

Junction tree notes

2009-03-16T15:11:00.003-04:00

I needed a brief and precise refresher for junction tree algorithms to review a paper. I found the following lecture notes from Mark Peshkin very useful:
http://ai.stanford.edu/~paskin/gm-short-course/lec3.pdf

getting cgi to work on Mac

2009-03-10T11:41:00.002-04:00

Sorry for this non-vision-ey, non-ML-ey post, but I need to make a note of this to save some time and energy and time in future. Who knows, you might find it useful someday.

Steps to follow:
1) httpd -V (to find out SERVER_CONFIG_FILE)

2) edit this file as following (would need root/super-user access):

(a) insert (somewhere)

<Directory "/Users/*/Sites/cgi-bin">
AllowOverride None
Options ExecCGI
Order allow,deny
Allow from all
</Directory>

(b) Uncomment

AddHandler cgi-script .cgi

You would need to turn Web Sharing OFF and then ON to bring the changes into effect.

SUnS'09

2009-02-02T23:37:00.003-05:00

I attended the scene understanding symposium at Boston. There were some very good talks (I missed two sessions which I definitely did not want to). I missed the first half of Jeremy Wolfe's talk on "Search in real scenes: The latest mysteries, the latest clues." It was interesting.

One interesting thing I learned from Michael Paradiso's talk was about some experiments (--Need link--) showing that the regions of brain that are involved in early stages of processing continue to be active in even the later stages. This suggests that a pipeline approach to model it may not be useful.

Uri Hasson had an entertaining talk on estimating temporal responses in the human brain while watching modified clips (sequential, reverse sequential, etc.) of Charlie Chaplin's videos.

I missed David Forsyth's talk. Ce Liu's talk about dense scene alignment was very impressive. He has code and results on this project webpage.

NIPS' paper list -- revisited/filtered with one-line detail.

2008-12-30T10:34:00.005-05:00

# Modeling the effects of memory on human online sentence processing with particle filters
R. Levy, F. Reali, T. Griffiths
-- incremental, limited-memory model for understanding sentences using particle filters.

# An ideal observer model of infant object perception
C. Kemp, F. Xu
-- perception is guided by the principle of persistence, i.e. things tend to remain the same and mostly follow rigid motion.

# A rational model of preference learning and choice prediction by children
C. Lucas, T. Griffiths, F. Xu, C. Fawcett
-- econometric model for explaining a young child's use of statistical information to infer preferences. very interesting.

----------------------------------------
# Bounds on marginal probability distributions
J. Mooij, H. Kappen
-- bound on single-variable marginal probability distributions in factor graphs by propagating bounds (convex sets of probability distributions) over a subtree of the factor graph, rooted in the variable of interest. Bounds its approximate Belief Propagation marginal, or belief, as well.

# Domain Adaptation with Multiple Sources
Y. Mansour, M. Mohri, A. Rostamizadeh
-- convex combination of multiple source hypothesis can perform poorly. there exists a distribution weighted combination that achieves the same error as the maximum error on all the sources. (this paper could throw some insight for distributed IR .. may be that is exactly this paper is about).

# Beyond Novelty Detection: Incongruent Events, when General and Specific Classifiers Disagree
D. Weinshall, H. Hermansky, A. Zweig, J. Luo, H. Jimison, F. Ohl, M. Pavel
-- HAVE to see. experiments include face recognition for audio+video data.

# Accelerating Bayesian Inference over Nonlinear Differential Equations with Gaussian Processes
B. Calderhead, M. Girolami, N. Lawrence
-- GP regression to accelerate inference.. should look in more details.

# Generative and Discriminative Learning with Unknown Labeling Bias
M. Dudik, S. Phillips
-- entropy-based weighting offers an improvement over constant estimates of class proportions, consistently reducing log loss on unbiased test data... not sure about the generalizability here.. need to look at the full paper.

NIPS papers by title-- my selection

2008-12-20T11:23:00.002-05:00

My selection (solely based on the title) grouped according to my inference about their content. I will look at the abstracts soon (hopefully), and then look at some of these more carefully.

----------
# Modeling the effects of memory on human online sentence processing with particle filters
R. Levy, F. Reali, T. Griffiths
# Analyzing human feature learning as nonparametric Bayesian inference
J. Austerweil, T. Griffiths
# An ideal observer model of infant object perception
C. Kemp, F. Xu
# A rational model of preference learning and choice prediction by children
C. Lucas, T. Griffiths, F. Xu, C. Fawcett

---------------
# One sketch for all: Theory and Application of Conditional Random Sampling
P. Li, K. Church, T. Hastie
# Bounds on marginal probability distributions
J. Mooij, H. Kappen
# Rademacher Complexity Bounds for Non-I.I.D. Processes
M. Mohri, A. Rostamizadeh
# Domain Adaptation with Multiple Sources
Y. Mansour, M. Mohri, A. Rostamizadeh
# Comparing model predictions of response bias and variance in cue combination
R. Natarajan, I. Murray, L. Shams, R. Zemel
# Beyond Novelty Detection: Incongruent Events, when General and Specific Classifiers Disagree
D. Weinshall, H. Hermansky, A. Zweig, J. Luo, H. Jimison, F. Ohl, M. Pavel
# Accelerating Bayesian Inference over Nonlinear Differential Equations with Gaussian Processes
B. Calderhead, M. Girolami, N. Lawrence
# Generative and Discriminative Learning with Unknown Labeling Bias
M. Dudik, S. Phillips

-----------
# Relative Performance Guarantees for Approximate Inference in Latent Dirichlet Allocation
I. Mukherjee, D. Blei
# Learning Taxonomies by Dependence Maximization
M. Blaschko, A. Gretton
# DiscLDA: Discriminative Learning for Dimensionality Reduction and Classification
S. Lacoste-Julien, F. Sha, M. Jordan
# Deflation Methods for Sparse PCA
L. Mackey
# Bayesian Exponential Family PCA
S. Mohamed, K. Heller, Z. Ghahramani

----------
# SDL: Supervised Dictionary Learning
J. Mairal, F. Bach, J. Ponce, G. Sapiro, A. Zisserman
# Cascaded Classification Models: Combining Models for Holistic Scene Understanding
G. Heitz, S. Gould, A. Saxena, D. Koller
# A "Shape Aware" Model for semi-supervised Learning of Objects and its Context
A. Gupta, J. Shi, L. Davis

Richard Hamming's talk on doing top quality research

2008-12-20T00:25:00.002-05:00

I stumbled on a transcript of Hamming's talk from 1986, where he talks about doing "Nobel-Prize research" as opposed to "good research". While his concerns about the role of computing appear to be unequivocally addressed in the current of science, the points he makes about how a good researcher could be more productive and successful, will always remain valid. Although the transcript is pretty long and incoherent in parts (IMHO), I found it to be one of the most inspirational, yet practical, talks about doing good research. Lots of big names (Shannon, Feynman, etc.) and related anecdotes/events are mentioned.

My favorite points: open doors, knowledge/productivity (or the lack of either) as compound interest, reading papers to learn about the problems (and not the solutions), appearance of conforming, and (for pretty much the first time by a researcher) acknowledgment of management as a reasonable choice.

LaTeX editor for windows

2008-12-14T22:26:00.003-05:00

Although I am happy with vim as a general editor, I prefer to use a specific editor for LaTeX (mostly for quick reference and selective compilation). In past, I have tried TeXnicCenter and WinEdt -- the latter remained my favorite for being lightweight and more integrated with MikTeX (from a usability perspective). I finally got annoyed by the incessant bugging message to register the product (as the trial version expired long ago). So I am trying LEd (http://www.latexeditor.org/ ) -- its looking pretty good so far. The interface is pretty messy though.

WEKA (and command line issues)

2008-11-18T10:02:00.003-05:00

Despite the incessant pestering by a lab-mate (Moe), I never got motivated to explore WEKA. For my recent "20%"-project on discriminative-generative debate, I thought I would look into WEKA. Within days (if not, hours), now I am totally loving it. A lot of useful ML components have been implemented in this toolkit. I wish I started using it a few years ago -- may be I could have had codes for some popular topic-models, including people-LDA, and random field, including SHRF, packages added to this toolkit, at the least.
(that reminds me of the unfortunate fate of my under-development C++ library that I called TOMOLIVE -- Topic models for learning in vision .. boohoo! )

Although the Weka's Experimenter GUI provides a lot of options to play with, I could not find the knob I wanted to turn, the one that iterates over the data split size. So I decided to use the command line. With multiple parameters to pass to the classifier classes that in turn are passed to the base class (and similar things happening with split evaluators), I got completely lost in the Unix-syntax. I finally got it to work in a not-so-elegant/non-extensible fashion. I want to do something like the following:

A x (B y (C d) )

or, the base class (that implements main) takes as argument x and B, where B is specified by options y and another class C (which is further specified by argument d).

As per my knowledge, the following two ways should implement this:
1. A x B -- y C -- d
2. A x "B y \" C d \" "

But, in practice, I do not observe the same result. Perhaps, there is something amiss about these expressions. I could tweak the first expression to work for me but this format falls apart when A takes another class E (with argument specifications) in addition to x and B, as "--" appears to force a single non-leaf child on the parsing of the command line. Any suggestions?

presentation tips ? (contd)

2008-11-11T21:22:00.002-05:00

Adding to my previous post:
* I wonder if there should be at least one "hard" slide or concept in the presentation, which only a few will get.

One common suggestion about job talks is that the presenter should include some stuff that is a little in-depth to catch attention of the people who are closely following the related line of research -- mostly, to demonstrate the hardness of the problem and the significance of the presented result. I believe similar kind of expectations can be assumed for any other presentation as well. When people say "no equations", it does not mean "absolutely no equations" -- it is rather "no equations in most of the presentation, and a couple of key equations on that `hard' part of the presentation". In other words, there should be some motivation for the audience to look at additional details after the talk. They should not leave with a feeling that they understood everything. There are definitely some negative aspects of this approach, and I have a feeling that only a few people will approve of this. But, in my opinion, this has practical implications and avoids the risk of the talk being perceived as simplistic.

Course on pattern recognition

2008-11-10T10:34:00.003-05:00

Following up on my previous post on introductory courses for learning in vision, I am adding the link for Chris's course on Advanced Topics in Pattern Recognition at University of Rochester.

Although this course has very little to do with computer vision, I think the syllabus outlines most of the machine learning tools prevalent in learning in vision. One thing I like about this course is its breadth. This course would probably not go into the intricacies of any of the concepts/models, but it is very likely to motivate students to explore further details on their own.

My take on CNN hologram

2008-11-08T11:56:00.005-05:00

Since every Tom, Dick, and Harry is talking about CNN hologram, I thought I would also write something about it (without parenthesizing the word hologram with quotation marks, that is). Clearly, there are two camps out there: one that is too amazed by the cool technology, and the other which thinks it was horrendous. When I saw it on TV, I oscillated between the two camps, but, eventually, unsubscribed from both of them.

The CNN hologram reminded me of the system that CMU deployed at Superbowl in 2001. Prof. Takeo Kanade gave a talk at CVPR'06 in NYC. I was totally amazed by learning the details of this system: hundreds of cameras installed all around the field, miles of wires, numerous computers -- all synchronized to produce the cool technology. A true engineering marvel! It might be a very far stretch to draw similarities between the two systems, and I am not trying to do that; it is just that CMU's system was the first thing that came to my mind. Anyways, as compared to this system, CNN hologram acquisition setting appears rather miniature in terms of the complexity of both the scene and the tracking issues. I am not sure if in the Superbowl system, the vision problems were circumvented by the over-constrained camera systems, but (to me) that definitely appears to be the case in the hologram system.

One technology that I would like to see in near future (and I believe the required positive research results exist) is better processing of audio signal, mostly when SNR is very low. In particular, when the journalists were reporting amidst the crowd, even though they were shouting at the top of their lungs, the signal was inaudible. I think, with the state-of-the-art audio processing techniques, this problem should be manageable.

I think what I wish to see in future is a clearer audio signal and less constrained settings for the hologram acquisition. In other words, I wish to see the person "beamed" directly from outdoors (as opposed to from inside a customized trailer), which I suppose will open a can of "vision-related" worms (problems).

That said, I am very excited to see the fast transition of the research in vision, signal processing, and human computer interface into mainstream / popular technology. No matter how hard the cynics lambaste/ridicule these demonstrations, there is no doubt that these things are inspiring a lot of young and creative minds.

Feed Reader

2008-10-30T09:15:00.002-04:00

I recently started using Google reader -- and the infatuation phase has so far lasted for almost a week now. I found the following: (1) new interesting blogs particularly the ones for the journals (bye bye mailing list.. beta phase :D ), and (2) I could actually keep track of the blogs synchronously with the changes (which is both good and bad). I still have not figured out a way to make it detect the comments on different blogs -- (as Gary suggested) it might only be done at the blog end and not at the feed-reader end. One last thing, I was surprised and motivated by the number of subscribers to this blog. YAY!

I have some stuff to post but I should be working on tomorrow's talk at University of Rochester :|

Intro to vision courses

2008-10-27T11:32:00.003-04:00

Many of my machine learning friends have asked me about some pointers for the intro to learning in vision kind of stuff. They are mostly looking for how some introductory courses that provide some outline and perhaps motivation for interesting problems in computer vision where machine learning techniques/methods were found to be useful.

In a recent reply to one such inquisitive friend (who asked a similar question about Bayesian methods in Vision), I wrote:

Object recognition is a very broad field with zillions of different approaches being used: statistical, probabilistic, and deterministic, among others. In particular, Bayesian methods have been used in computer vision and pattern recognition for perhaps four-five decades now. Many people would agree on partitioning the approaches used for vision as: model-based vision, and learning-based vision. The boundary between the two is a little fuzzy though. That said, I would add that computer vision is a practical application domain for many of the machine learning techniques developed so far (in fact, many of them including Gibbs sampling) were originally developed for vision tasks only. Thus, it is expected to see variants of machine learning / statistical techniques appearing in vision literature as well.

Following on this thread, she asked me about any relevant surveys/courses that could be helpful. This question made me realize (duh!) that most of the intro to vision courses focus mostly on describing the basic framework (imaging, color, etc.) and (often) common problems such as segmentation, and the primitive ways of solving them, giving an impression of computer vision as a stagnant, or at least very slow moving, research field. For instance, most of theses courses would stop at Canny edge detectors for semantic grouping of line-like features and Eigenfaces for face recognition, which (IMHO) does justice neither to the complexity of these problems nor the magnitude of efforts put towards solving them.

In my subsequent reply, I wrote:

The standard computer vision courses focus more on the basic framework such as details of imaging and colors, which might not be very interesting. I guess you would be better off looking at advanced courses related to "advanced perception", "pattern recognition", or "image understanding.

The following might be useful pointers:
Alyosha Efros's course at CMU
Antonio Torralba's at MIT
David Lowe's at UBC
Jitendra Malik's at UCB
David Forsyth's at UIUC

Interesting papers at ECCV'08

2008-10-19T11:56:00.004-04:00

Although I was not at ECCV, my good old friend Professor Weinman was there (His pendulum was pointing East :D ). He shared his list of papers-to-see with me.

* Learning to Localize Objects with Structured Output Regression (Best student paper)
* Learning Spatial Context: Using Stuff to Find Things (Best paper) -- see previous post on papers from ECCV
* The Naked Truth: Estimating Body Shape Under Clothing (catchy :) )
* Improving People Search Using Query Expansions: How Friends Help To Find People (related to People-LDA) -- see previous post on papers from ECCV
* SIFT Flow: Dense Correspondence across Different Scenes (flashiest and most interesting paper, at least from "neat-o" standpoint)
* Discriminative Sparse Image Models for Class-Specific Edge Detection and Image Interpretation (Training class specific Pb (edge detector) to help object detection)
* Learning Optical Flow (Hail MRF!)
* Beyond Loose LP-relaxations: Optimizing MRFs by Repairing Cycles (better optimization than graph cuts, sincere request to the authors to make the code available)

Thanks Jerod!

Why MCMC works?

2008-09-14T14:24:00.011-04:00

Long ago, I gave a chalk-talk for the theoretical details of why MCMC works? I apologize for the lost references; these notes are mostly extracted out of online tutorials, classroom notes, and some books. These notes are more like cheat-sheet, and may not be very readable. Nevertheless, the step-by-step progression could provide useful pointers.

Importance Sampling

E_p[f] = \Int_X p(x) f(x) dx = \Int_X (p/q * f) * q dx
= \Int_X g(x) * q(x) dx
= E_q [p(x)/q(x) * f(x) ]
- unbiased, but has high variance.
To make it low variance, use the normalization, commonly called weighted importance sampling. The actual formulation for weighted sampling is motivated by the fact that we can not compute p(x)/q(x), rather we can only compute p*(x)/q*(x), where f* represents un-normalized distribution, and f~ denotes sample distribution for f. Note that the new estimator is biased.
We refer to p(x) as target and q(x) as proposal distributions. q(x) is a distribution which is easy to sample from, e.g., uniform, Gaussian, Cauchy, exponential).

Rejection Sampling

- generate y ~ q(.)
- sample u ~ Unif[0, c*q*(y)]
- if u > p~(y), reject y
~~~~ otherwise, accept y
p_sample(y) = p(y)
p_accept = Area(accept) / Area(accept) + Area(reject) \prop 1/c

Metropolis-Hastings algorithm

Given a target distribution p(x) = p*(x) / Z,
- Initialize X(0) arbitrarily
- For iteration t = 0, 1, 2
~~~~ > draw y from proposal distribution, q(.| X(t)), call it the transition function T(X(t), .))
~~~~ > accept/reject: sample u ~ Unif[0,1]
~~~~ > X(t+1) = ~~ y, if u \lt A(X(t), Y),
~~~~ ~~~~~~~~~~~~~ X(t), otherwise,
~~~~ ~~~~~~~ where A(X(t), Y) = min (1, p*(y)T(y, x(t)) / p*(x(t))T(x(t), y) )

Special case

T(x,y) = T(y,x) i.e., T is symmetric -- Metropolis algorithm, in which case,
A(x(t), y) = min(1, p*(y)/p*(x(t)) ). Thus, if p~(y) \geq p~(x(t)), always accept, otherwise accept with probability p~(y)/p~(x(t)) ( < 1)
Analogy: Always go uphill, and go downhill with some probability, so we are not stuck in a local maximum.
In a way, this algorithm is performing a random walk in the entire space.

Background on Markov Chains

- X(1) ... X(t)
- X(0) ~ q_0(.)
- transition X(t) --> X(t+1), specified by S(X(t), X(t+1)) = q(X(t+1) | X(t)). note that \Sum_y S(x,y) = 1.
- assume homogeneous chain, i.e., S is the same for all t.

Working

- initial distribution, q_0
- compute q_(t+1)(y) = \Sum_x q_t(x) S(x, y) {matrix notation: q_(t+1)' = q_t' * S}
Note that S is row-stochastic, i.e. S*1 = 1

Why it works?

We want to know if q_t converges to some limit, and if does, what does it converge to?
Definition: We say that \pi \in R^k is invariant (w.r.t. S) if \pi \geq 0, 1'\pi = \pi, and \pi' * S = \pi. Alternatively, \pi is a fixed point for the update equation.
Idea: Let X(0) ~ \pi, then q_t = q_0 * S^t = \pi* S^t = \pi. In other words, if we start in \pi, we stay in \pi. Hence for convergence, we want to reach such a point where we have q_t as invariant (w.r.t. S).
Note that the matrix S \in R^(k x k) is (a) non-negative, i.e., S(i,j) \geq 0, and (b) row-stochastic, i.e., S*1 = 1.

Perron-Frobenius theorem

Definition: spectral radius of a matrix S is defined as the highest absolute eigenvalue of the matrix.
Every non-negative, row-stochastic matrix S satisfies \rho(S) = 1. Let the corresponding left eigen-vector be v. v \gt 0, and v'S = v. So we can set \pi (in the above discussion) to the vector \pi_i = v_i / (\Sum_j v_j).
Corollary: Every finite-state Markov chain has an invariant distribution.
So we conclude that q_t converges. Next question: is it unique? In other words, we want to find out conditions when q_t converges to a unique \pi (now that we know that such a limit exists).
A related question is: when is an invariant distribution unique?
Ans: when the Markov chain is ergodic.
If the Markov chain is ergodic and aperiodic, then there is a unique invariant.

Back to Metropolis-Hastings

First, check if \Sum_x p(x) S(x,y) = p(y).
Remember, S(x,y) = T(x,y) min (1, p*(y) T(y,x) / p*(x) T(x,y) )
Definition: A matrix S satisfies detailed-balance w.r.t. \pi iff
\forall x,y, \pi(x)S(x,y) = \pi(y)S(y,x).
(also called reversibility, i.e., moving in both direction in the chain are equivalent)
Lemma: Any such \pi is invariant w.r.t. S.
So it suffices to check that S(x,y) is in detailed-balance w.r.t. the target distribution, which is true. Thus p(.) is invariant w.r.t. S.

Gibbs Sampling

To make it more practical and easy to handle, we need careful ways to design S.
Gibbs Sampling is one such special type of Metropolis-Hastings algorithm.
- transitions easy to draw from.
- Consider random vector (X1 ... Xn), where n is the number of nodes in the graphical model.
~~ initialize X(0) = X1(0) ... Xn(0) arbitrarily.
~~ For iteration t = 0,1,2...
~~~~~~~ update Xi(t+1) ~ p(.|X1(t) ...Xi-1(t) Xi+1(t) ... Xn(t))
i.e., walk through the graph and update using local conditional probabilities.

Product of Experts and Contrastive Divergence

2008-09-04T20:20:00.002-04:00

So we are LIVING again.

Instead of posting my notes here, I am posting a pointer/reminder to check the entries for "Boltzmann machines" and "product of experts" on scholarpedia.

Face data sets

2008-08-29T13:31:00.003-04:00

Lior Shamir recently made a very clever (shocking?) observation (IJCV preprint) about many of the existing face data sets, which makes a strong argument against the validity of many of the published empirical comparisons for face recognition task.

The main idea is that for many of the popular face data sets, the classification accuracy using a small non-facial area in the images is much higher (and sometimes in 90+ percentages) than the corresponding random accuracy. For example, they obtained a non-facial accuracy of 99% for Yale B data set.

Personally, it was surprising for me to see that Indian Face Database fared better than other data sets in terms of this "hard"-ness property, except for Color FERET, which has a large number of different people in it but AFAIK it is not publicly/freely available for download.

Two research questions

2008-08-11T10:39:00.003-04:00

I am wondering what the answers to the following questions are. I think I can work out the first one (I know it is not ANOVA), at least its first part (a). The second is more interesting, and, perhaps, could be more widely applicable.

1) Expected variance
-----------------
Given a set S and its subsets T_i (i = 1 ... k), and the corresponding variances var(S) and var(T_i), for a random subset T of S, what is p(var(T) >= Theta) assuming
(a) x \in S is distributed according to a Gaussian distribution (\mu, \Sigma)?
(b) non-parametric distribution following the observations.

2) Max-min analogue of integral images:
------------------------------------
Find the minimum integer value of k such that k-max chains at all pixel locations are sufficient for an O(k) time determination of a "random-access" subset max (or min)?

I could not find anything in the literature that is relevant to these. I would appreciate if

A nerdy joke...

2008-08-07T17:18:00.005-04:00

In response to the Things and Stuff model by Heitz and Koller at ECCV08, I thought I would make a more useful model ;) (jk!)

THE graphical model for life, universe, and everything (Yes! it is directed and acyclic) (Copy- right and left: Vidit Jain.. muhahaha!).

A few notes:
1. The subset of variables that is observed depends on the observer.
2. Believe it or not! N, M, and K are fixed.
3. Y is a binomial random variable with parameter known to a few.
4. Depending on your faith, there can be a plate on "GOD" variable.
5. GODs may also interact among themselves.
6. There is no known evidence of "real" co-operation among MANs.
7. And of course, the model parameters are not shown in this graphical model for
a reason. (HINT: 42)

Happy MCMC-ing!!!

Some papers from ECCV'08

2008-08-07T15:57:00.004-04:00

* Mensink et al. Improving people search using query expansions: How friends help to find people.
-- Cool idea for expanding visual people by expanding the query from "A" to "A -B -C". In other words, look in the database for people co-occurring with the queried person for candidates to discriminate against while making a correspondence for a face in the image with the queried name.

* Heitz and Koller. Learning Spatial Context: Using Stuff to Find Things.
-- Nice demonstration of the utility of the stuff-things relationship in a generative framework. The modeling of a latent variable corresponding to "stuff" for image-regions is similar to Selective Hidden Random Field framework (one of the obvious differences is Bayesian Network vs. Markov Network framework). Learning of selection of features is also interesting, however, a Markov network appears to be more suitable and straightforward choice for this type of modeling.

Checklist for making a "good" computer vision presentation

2008-08-03T18:32:00.002-04:00

From my past experience at watching really good, enjoyable presentations -- such as Shree Nayar's talks -- I learned a few simple tips to improve my own presentations. I made an incomplete, in-progress checklist that I refer to while preparing a new one. I posted it on my web-site:

http://vis-www.cs.umass.edu/~vidit/visionTalkTips.html

Feel free to send in comments to add to it.

UMass's loss is Grinnell's gain

2008-07-10T07:55:00.002-04:00

The transition from Mr. Weinman to Dr. Jerod Weinman will see a significant change in the future vision meetings. The good news is now I (and other presenters) can get away with a lot of hand-waving when it comes to talking about inference. The bad news is I have lost my O'Reilly reference books from the shelf next door (cube).

Good luck to Jerod.

CVPR08: papers to peruse/skim through

2008-07-03T16:27:00.003-04:00

I missed some of the sessions on Tuesday and wednesday afternoons, so I have to look back at those to see what was interesting there. Same is the case with the workshop papers. Most of the handwritten notes in the tiny CVPR booklet are too illegible to ignite the non-Bayesian, nonparametric inference in the neural network residing in my brain.

* He and Zemel. Latent Topic Random Fields...
* Todorovic and Ahuja. Learning subcategory relevances to category recognition
* Han et al. Discovering Class specific compositive features through discriminative sampling with Swendsen-Wang Cut.
* Zeng and Van Gool. Multi-label image segmentation via point-wise repetition.
* Ramalingam et al. Exact inference in multi-label CRFs with Higher order cliques.
* Alahari et al. Reduce, Reuse, and Recycle: efficiently solving multi-label MRFs.
* Park et al. Efficient mean shift belief propagation for vision tracking.
* Ramnath et al. Increasing the density of active appearance models.
* De la Torre and Minh. Parametrized KPCA: Theory and application to image alignment.
* Parag et al. Boosting Adaptivelinear weak classifiers for online learning and tracking.
* Xu and Roy-Chowdhury. A theoretical analysis of linear and multi-linear models of image appearance.
* Meeds et al. Learning stick figure models using nonparametric Bayesian estimation. (could possibly be extended for multiple animals in a scene to give an "Alaska Zoo Model" :D )
* Ni et al. Epitomic location recognition.
* Subbarao et al. Robust unambiguous parameterization of the essential manifold.
* Wang et al. Manifold-manifold distance with application to face recognition based on image set.
* Thurao and Hlavac. Pose primitive based human action recog. in video or still images.
* Desealers et al. Pan, zoom, scan - time-coherent, trained automatic video cropping.
(Also look at CVPR07 paper by Huo et al about robustt feature computation in 12 lines of code :D ) -- cool device-specific video adaptation
* Wu and Nevatia. Optimizing discrimination-efficiency tradeoff.. for object detection.
* Wu et al. Segmentation of multiple, partially occluded objects...
* Kokkinos and Yuille. Scale invariance without scale selection.
* Fritz and Schiele. Decomposition, discovery, and detection of visual categories using topic models. -- LDA on HoG features actually WORKS !!
* Jain et al. Fast image search for learned metrics.
* Hong Cheng, Zicheng Liu, Nanning Zheng and Jie Yang. A Deformable Local Image Descriptor

Short papers at conferences: revisited

2008-06-30T17:33:00.003-04:00

Revisiting this topic discussed in a previous post:
http://vimsu99.blogspot.com/2008/02/short-papers-at-vision-conferences.html

From talking to some people at CVPR about the possibility of having short papers at vision conferences, I learned a few interesting things. Although not new or for the first time, I am pretending that I was not even thinking about these issues.

Apparently, most of the vision researchers agree that there is no clear distinction between the "oral" papers and "poster" papers; the cut-off is fairly arbitrary and has nothing to do with an analysis of the "quality-curve" of the accepted papers. This phenomena also holds true (to some extent) about the best paper award as there is usually a group of deserving papers and not a clear winner. However best paper award can be left out from this discussion as it still implies a higher percentile over the accepted papers.

The "oral" status has thus artificially been elevated above the "poster" papers. Despite the fact that everybody accepts this case, it is unfortunate (for obvious reasons) that everybody is implicitly giving more importance to orals (through biased promotion and perception, i.e., "he/she got three oral presentations"). So the concern is that if short papers come into existence, there would be an additional synthetic division in the perception of the quality of conference papers.

I disagree with this reasoning because of the following two reasons:

(1) The oral-poster partition is due to the reviewer evaluation and not as authors' projection. In other words, the authors are often the best suited evaluators of their own work, so they should make a preliminary judgment of the possible impact. If they get a very encouraging response at the conference, they can make a journal/extended version with additional details.

(2) An introduction of short papers may perhaps reduce the reviewing time per paper as unnecessary details could be avoided (which are otherwise prevalent to avoid yet another unfortunate perception, "a six- (or fewer) page paper is not worthy of acceptance"), which is an incentive for the authors as well.

Once again these two points are again on the idealist frontier with a clear negligence towards the rat-race of publishing. If that was not the case, there would not have been any problem with the original format, in the first place. It is conceivable that short papers would rather be detrimental to the optimization of the number of publications (once again there would be a difference in the perception of the weights of these types of papers) required for the recipe of survival in current academic environment.

CVPR 08: reporting from Anchorage

2008-06-27T15:44:00.002-04:00

This report has been delayed because of the really weak internet connection at the conference venue and at the hotel where I am staying. The unfortunate part is that because of this delay, most of the details are wiped from the memory cache. So I guess I have to go back to the proceedings to actually make any intelligent remarks. Nevertheless, the scribblings in my small booklet could be helpful in shortlisting the otherwise huge list.

Just one comment, I am super-tired. The poster sessions were too crowded and stuffed, with very little time for the discussions to be fruitful, effective, efficient, or at the least enjoyable.

Also, just a request for the readers of this blog that I met at the conference (and others also): I would really appreciate if you can send me a link of your research blogs (if you have one) or post your pick of papers / comments / reviews from CVPR on this blog.

Is learning really helpful for evolution?

2008-05-07T10:59:00.002-04:00

In a recent article in NY Times titled "Lots of Animals Learn, but Smarter Isn’t Better", some interesting results from the research of Dr. Tadeusz Kawecki's group at University of Fribourg and Dr. Reuven Dukas's group at McMaster University is reported. This article discussed the trade-off between the benefits and costs associated with learning in fruit flies. The researchers cleverly designed experiment to stimulate learning in a selected group of fruit flies. After a few generations, although these tiny creatures became "smarter" (fast learners), their survival capabilities became inferior compared to the "non-smart" flies.

In the concluding remarks, these scientists hypothesize that (a) "learning evolves to higher levels only when it is a better way to respond to the environment than relying on automatic responses," and (b) "Forming neuron connections may cause harmful side effects", (c) "Each species evolves until it reaches an equilibrium between the costs and benefits of learning."

In the context of evolution of intelligence in humans, Kawecki, perhaps light-heartedly, posits that one just has to be smarter than others to survive or make progress.

some learning theory: generalizability

2008-04-30T17:36:00.003-04:00

I came across this super-awesome paper: Poggio et al. General Conditions for Predictivity in Learning Theory, Nature, March 2004.

I am so glad for a verification of many of my concerns with the classical learning theory. And of course, the results included in this letter to Nature are interesting, motivating, and surprising.

Here are my notes:

Notation
-----------
* Training: synthesizing a function that best represents the relation between the inputs x_i and the corresponding outputs y_i
* ERM (Empirical risk minimization) : the algorithm looks at the training set S, and selects as the estimated function the one that minimizes the empirical (training) error over the functions contained in a hypothesis space of candidate functions.
* Consistency: the expected error of the soluction converges to the expected error of the most accrate function in the hypothesis class.
* CV_{loo} (Cross-validation leave-one-out) stability measures the difference in errors at a point z_i between a function obtained given the entire training set and one obtained given the same training set but with the point z_i left out.
* CVEEE_{loo} : slight variant of CV_{loo} that includes the stabilty of empirical error and the stability of expected error.

Results
---------
* Classical learning theory (such as VC-theory) completely characterizes the necessary and sufficient conditions for generalization of ERM and its consistency.
* For ERM, generalization is equivalent to consistency.
* CVEEE_{loo} stability is sufficient for generalization for any learning algorithm (CV_{loo} is not, though) and necessary and sufficient for generalization and consistency of ERM.
* From VC-theory: simple theories should be preferred among all of those that fit the data.
* CVEEE_{loo} stability conditions correspond to the statement that the process of research should only incrementally change existing scientific theories as new data become available.

Etymology of common learning-related words such as recognize

2008-04-28T12:36:00.004-04:00

Being a non-native English speaker, I am sometimes confused about the precise use of words in a technical sense. In particular, when people talk about face recognition and face identification, it takes some time to build a common ground. Sparked by this annoyance and motivated by a paper from CVPR'08 -- Malisiewicz and Efros. "Recognition by Association via Learning Per-exemplar Distances" -- led me to look into the origin of these words to clear this confusion. As is expected from the English language, the issues still remain unresolved. Nevertheless, I learned some subtleties. Here are my notes:

* association: join with
* categorize: to affix a label to
* classify: putting into one of the fixed, known classes
* correspondence: harmony, agreement
* detect: uncover, disclose.
* identify: regard as the same
* recognize == re-cognize : to know again, perceive something or someone as already known

The most confusing pairs would be categorize & classify, and identify & recognize.
For the first comparison, I concluded that categorization could imply an infinite, and perhaps, unknown set of labels, whereas classification is associated with the division of the sample space into fixed classes. Also, categorization can be hierarchical but classification should not be.

I am still not sure about the ambiguity between identification and recognition. However, I believe that identification has do with the "sameness" of the observed signal. Recognition, on the other hand, is concerned with the perception. For example, while listening to a song on a radio, you could identify the song, but when you are watching the video of the same song without the audio signal, you would be recognizing the song (you could simultaneously identify the video, though).

Discussion on Generative vs. Discriminative models

2008-04-21T12:01:00.005-04:00

I recently led a discussion on generative vs. discriminative models in the context of research in IR in a seminar organized by Prof. James Allan. I am not sure how long the wiki for that course would be alive, so I am copying the stuff onto this blog.

Background
This seminar would focus on the comparison and contrast between two popular paradigms of machine learning: generative models and discriminative models. We will discuss some of the theoretical results relevant for this debate, focusing primarily on identifying the characteristics of the problem domain suitable for these methods. We will also look into some of the recent observations made in the fields of Information Retrieval and Computer Vision.

Required papers
* A. Ng, M.I. Jordan. On Discriminative vs. Generative classifiers: A comparison of logistic regression and naive Bayes. NIPS 2001 -- presents interesting theoretical comparison and results for the two paradigms. Just focus on the understanding of the theorem/lemma statements; the proofs are inconsequential for this discussion.
* R. Nallapati. Discriminative Models for Information Retrieval, ACM-SIGIR, 2004 -- a discriminative model for IR.
* J. Lafferty and C. Zhai. Probabilistic relevance models based on document and query generation, Workshop on language modeling and Information Retrieval, 2001. -- a generative model for IR. (The underlying assumption here is that everyone has already read Ponte and Croft's seminal paper on language models).

Required skims
Please skim through the Abstract, Experiments(Tables/Figures), and Discussion/Conclusion sections to get a feel of these computer vision papers. We will discuss aspects that are relevant to the generative-discriminative debate.
* V. Jain, A. Singhal, J. Luo. Selective hidden random fields: Exploiting domain specific saliency for event classification, CVPR 2008. -- a discriminative model.
* V. Jain, E. Learned-Miller, A. McCallum?. People-LDA : Anchoring topics to people using face recognition, ICCV 2007. -- a generative model.

Recommended
* X. Wei, W.B. Croft, Investigating Retrieval Performance with Manually-Built Topic Models, RIAO 2007. -- generative topic model.
* Y. Cao et al. Adapting ranking SVM to document retrieval, SIGIR, 06. -- discriminative.
* R. Raina et al. Classification with Hybrid Generative/Discriminative Models, NIPS 2004. -- best of both the worlds ?

Tangential but interesting
* D. Metzler. Direct Maximization of Rank-based Metrics. Technical Report, CIIR. -- presents a direct optimization for ranking as opposed to an indirect optimization based upon maximizing training data likelihood or classification accuracy.
* T.S. Jaakkola and D. Haussler. Exploiting generative models in discriminative classifiers. NIPS 98 -- deriving kernel functions for use in discriminative models from generative probability models.
* P.M. Long, R.A. Servedillo. Discriminative Learning can succeed when generative learning fails. COLT 2006. -- for theory-oriented minds.
* I. Ulusoy, C.M. Bishop. Generative versus discriminative methods for object recognition, CVPR 2005.

Questions
* Ng & Jordan suggest that we need 0(log(n)) training samples for generative models, and O(n) training samples for discriminative models to work well, where n is the dimension of the feature space. Since feature space to consider in IR is large (O(number of unique terms)), does it imply that generative models are more suited for IR tasks unless we have a very large training set (which we usually do not)? (Michael)
* Concerning the OOV problem: Given a particular task such as sentence retrieval or passage retrieval, say the best way to define the features for a discriminative model is over the words. So typically we will have to tackle the OOV problem by smoothing. How would this presumably affect the success of discriminative models (as opposed to the generative ones) for such IR tasks? (Elif)
* It seems to me there's an underlying question, which is whether IR is actually a classification problem or not---it is often /modeled/ as a classification problem, but is "relevant" really a class? I thought it was interesting in Ramesh's paper that he argues that the classification view models "the real-life IR problem very accurately" but then says "a user is primarily concerned about /how relevant/ a given document is". These two statements seem kind of contradictory to me. In my opinion---and this is intended for discussion---if "how relevant" is what you care about, you need to treat it as a ranking problem rather than a classification problem. (Ben)
* Would it be possible to overcome the problem of slower convergence towards the theoretical lower bound for discriminative classifiers by adding a time-dependent learning parameter? An approach like temporal-difference learning tends to put higher weight on the earlier training examples, and then "cooldown" as the number of examples increases. Intuitively, this seems like it may address the problem of needing more samples. An alternative could be a sort of "oversampling weight", where the earlier examples are actually treated as a number of examples, and subsequent examples (this would be in a separate epoch) would count as less, until you reach a 1-1 ratio of examples to "multiplied examples". (Marc)

Ng & Jordan, 2001 in simple words

2008-04-20T17:20:00.003-04:00

I have read Ng & Jordan's paper On discriminative and generative classifiers: A comparison of logistic regression and naive Bayes many times, and at every attempt ended up wasting some time struggling with the notation. So finally, I am penning down my interpretations of all of the results mentioned in this paper.

Notation: m = number of samples, n = dimension of the feature space.

Conclusion: The asymptotic error of the discriminative model is lower than the asymptotic error of the corresponding generative model (from the generative - discriminative pair), but the generative model require number of examples logarithmic in the dimensionality of the feature space to catch up with this asymptotic error (as opposed to linear in the case of discriminative models).

Prop. 1: Asymptotic error for discriminative (logistic regression) classifier is at most the asymptotic error for the generative (naive Bayes) classifier.

Prop. 2: For n-dimensional logistic regression, the error lies within O(sqrt(n/m log m/n)) of the asymptotic error.

Lemma 3: The parameters of the (naive Bayes) generative model converge to those of the asymptotic classifier with logarithmic number of samples, with high probability.

Theorem 4: The error also converges with a margin of error G(t), which in turn is upper-bounded by Pr(l_gen(x) \in [-tn, tn]).

Prop. 5: As long as most of the features are relevant to the class label, the expexted value of |l_gen,\infty(x)| will be \Omega(n)

Corollary 6: Under specific assumptions, the error of h_gen converges to h_gen, \infty with high probability when m = \Omega(log n).

selective hidden random fields

2008-03-08T14:08:00.008-05:00

Woohoo! I got a paper accepted at CVPR'08.

The title of this paper is:
"Selective Hidden Random Fields: Exploiting Domain Specific Saliency for Event Classification". The title is slightly longer than what I would prefer. Anyways, this paper introduces a specific sub-class of hidden-state conditional random fields that performs a selection of features simultaneously with classification. For a computer vision task, it corresponds to a joint segmentation-and-classification. This model is shown to be effective on sports images where it segments the playing surface in those images and uses them for classifying the sporting event.

While I am working on the camera-ready version, a draft is available at
http://vis-www.cs.umass.edu/~vidit/publications/cvpr08shrf.pdf.

Short papers at vision conferences

2008-02-13T13:40:00.003-05:00

Recently, John Langford wrote about complexity illness on his blog. He raised concerns about the growing bias for long and dense papers in machine learning. This observation is not novel in any sense, nevertheless, it is a serious concern. Computer vision is one of the worst affected fields by this phenomenon.

The number of submissions at CVPR, ICCV, and ECCV is way too high to project a honest reflection of the state of research. I am reviewing a couple of these. Most of the papers take up the maximum permitted space, i.e., eight pages, when the recommended number of pages is six. Some of these papers present incremental approaches. These papers often provide interesting insights into a particular problem or domain, but not enough to be accepted for publication (in my opinion). It would be much better if there is an option for short papers at the vision conferences.

Short papers are encouraged in conferences in various disciplines of computer science such as information retrieval, computational linguistics, and networks (I believe). These papers provide the authors an opportunity to present their research and achievements in a concise fashion, along with a judgment about their own work. These (short) papers would not be competing with the full papers (possibly talks) for acceptance. Moreover, the authors would not have to spend any effort to "prepare" the paper to counter the prevailing complexity bias.

Such categorization of papers, as full papers and short papers, may also help alleviate the reviewing situation. Currently, each reviewer is assigned more than ten papers for each of these vision conferences. It is becoming harder to spend enough time on all of these papers for helpful, constructive, and appropriate reviews, especially when many of these papers are filling in the entire space with unnecessary details and uninteresting diversions. An alternative of short papers would provide a mechanism for the authors to be more responsible to the community, and thus make these conferences more reflective of the true state of research in computer vision.

SUnS'08

2008-02-06T00:51:00.000-05:00

I attended the scene understanding symposium (SUnS'08) at Boston last Friday. I expected it to be very related to neuro-science and cognitive science. On the contrary, there were many machine vision talks and posters. Josef Sivic's presentation on walking through a collection of images in a fashion inspired by boundary extension phenomenon, was very impressive. Irving Biederman's talk was both entertaining and informative. I missed a couple of talks when I was talking with people interested in my poster.

Among the learning in vision posters I found the posters presented by Ariadna Quattoni (MIT), Kate Saenko (MIT), and Nicholas (U of Rochester) interesting.

It was also nice meeting the vision lab alumni (Dima and Frank). I had a great time in Boston.

Visual Objects in Context

2008-01-21T12:39:00.000-05:00

I read this amazing review by Moshe Bar on the contextual processing for object recognition. Here are my notes on the non-neuro part of this review.

---------------------------------------------------------------------------------
Bar, M. Visual Objects in Context Nature Reviews: Neuroscience, 2004, 5, 617-629

An excellent survey (review) of the cognitive psychology, neuroscience, and a few computational studies on using the context for object recognition.

Terminology:
1) Priming: An experience-based facilitation in perceiving a physical stimulus. In a typical object priming experiment, subjects are presented with stimuli (the primes) and their performance in object naming is recorded. Subsequently, subjects are presented with eiher the same stimuli or stimuli that have some defined relationship to the primes. any stimulus-specific difference in performance is taken as a measure of priming.
2) Context frames: According to a popular proposal for the contextual representation of associated objects (for cortical processing), the prototypical contexts are represented as structures that integrate info about the identity of the objects that are most likely to appear in a specific scene with info about their relationships. These structures are referred to as context frames. Also called schemata, scripts, and frames.
3) Boundary extension: A type of memory distortion in which observers report having seen not only information that was physically present in a pcture, but also information that they have extrapolated outside the scene's boundaries.

Five types of relationships characterizing a scene (Biederman) : (1) Support (physically supported vs. floating; (2) Interposition (e.g., occlusion); (3) Probability (some objects are more likely than others); (4) Position (typical location of objects); (5) Size (familiar relative size of objects).

- Visual objects are contextually related if they tend to co-occur in our environment.
- Representation in brain (hypothesis) : Grouping by objects in the occipital visual cortex, by basic-level categories in the anterior temporal cortex, by contextual relations in the parahippocampal cortex, and by semantic relations in the prefrontal cortex. In addition there can be other stored relations. There is one centralized, detailed object representation that serves all these relations 'on demand'.
- Isolated objects may still be easier to recognize than the object embedded in a contextually coherent scene because of possible difficulties with (a) segmentation, and (b) attentional distraction.
- When does context get processed? Competing hypotheses : (1) So rapidly that it facilitates perceptual analysis of objects; (2) when a context frame is activated it might sensitize the representation of objects associated with that context; (3) object recognition and contextual scene analysis are functionally separate and interact only at a later stage.

From a computational standpoint, it is clear that contextual representations may provide efficient generalizations in new situations and shortcuts in perceptual analysis.

vision and machine learning

2008-01-01T13:52:00.000-05:00

I was thinking about the dominant practices in computer vision. It is very unpleasant to acknowledge some of the usual trends. The worst of these are the ones related to adoption/application of existing machine learning techniques. There are, undoubtedly, many contributions in machine learning from "vision researchers", and there are numerous useful applications of learning algorithms to solve vision related problems. The point of concern is the growing negligence of the "old school" vision approaches such as physics and geometry based vision. The idea of solving real problems is becoming obsolete; building models is the new fashion statement. Is this the standard dilemma of systems versus algorithms? I think both of these perspectives are good ideally, but the implementations are not perfect: we end up with non-generalizable systems (even worse, with non-public implementations) and incremental algorithms. How are these useful for the community? Other than recognition (in the form of publications), what purpose are these papers serving? This raises a question: what is the distribution of the number of citations of papers from conferences like CVPR, ICCV and ECCV? A useful script to work on.

CVPR deadline

2007-12-14T14:32:00.000-05:00

I am so glad that the CVPR deadline was postponed. My paper improved a lot in that "extra" week. And now after a post deadline slump of two days, and finishing off some writing that was due today, I feel so relieved. Where are the forgotten lists of papers? Well, this slump can go for another couple of days... a Stephen King novel would not be bad either :D .

interesting papers from NIPS07

2007-11-24T10:59:00.000-05:00

# J. Ye, Z. Zhao, M. Wu: Discriminative K-means for Clustering
# D. Newman, A. Asuncion, P. Smyth, M. Welling: Distributed Inference for Latent Dirichlet Allocation
# J. Verbeek, B. Triggs: Scene Segmentation with CRFs Learned from Partially Labeled Images
# M. Mahdaviani, T. Choudhury: Fast and Scalable Training of Semi-Supervised CRFs with Application to Activity Recognition
# J. Bradley, R. Schapire: FilterBoost: Regression and Classification on Large Datasets
# S. Boutemedjet, D. Ziou, N. Bouguila: A Unified Model for Content Based Image Suggestion and Feature Selection
# N. Le Roux, Y. Bengio, P. Lamblin, M. Joliveau, B. Kegl: Learning the 2-D Topology of Images
# V. Ferrari, A. Zisserman: Learning Visual Attributes
# B. Russell, A. Torralba, C. Liu, R. Fergus, W. Freeman: Object Recognition by Scene Alignment
# P. Berkes, R. Turner, M. Sahani: On Sparsity and Overcompleteness in Image Models
# X. Wang, E. Grimson: Spatial Latent Dirichlet Allocation
# D. Blei, J. McAuliffe: Supervised Topic Models
# Y. Chen, L. Zhu, C. Lin, A. Yuille, H. Zhang: Rapid Inference on a novel AND/OR graph: Detection, Segmentation and Parsing of Articulated Deformable Objects in Cluttered Backgrounds
# K. Sinha, M. Belkin: The Value of Labeled and Unlabeled Examples when the Model is Imperfect
# Z. Lu, M. Carreira-Perpinan, C. Sminchisescu: People Tracking with the Laplacian Eigenmaps Latent Variable Model
# F. Richardson, W. Campbell: Discriminative Keyword Selection Using Support Vector Machines
# A. Chechetka, C. Guestrin: Efficient Principled Learning of Thin Junction Trees
# X. Nguyen, M. Wainwright, M. Jordan: Estimating divergence functionals and the likelihood ratio by penalized convex risk minimization

To read (ICCV 07)

2007-09-04T10:30:00.000-04:00

* Tiberio Caetano, Li Cheng, Quoc Le, Alex Smola: Learning Graph Matching
* Dashan Gao, Nuno Vasconcelos: Bottom-up saliency is a discriminant process
* David Liu, Tsuhan Chen: Unsupervised Image Categorization and Object Localization using Topic Models and Correspondences between Images
* Johannes Stallkamp, Hazim Kemal Ekenel, Rainer Stiefelhagen: Video-based Face Recognition on Real-World Data
* Krystian Mikolajczyk, Jiri Matas: Improving SIFT for Fast Tree Matching by Optimal Linear Projection
* Li-Jia Li, Li Fei-Fei: What, where and who? Classifying events by scene and object recognition
* Michael Jamieson, Sven Dickinson, Suzanne Stevenson, Afsaneh Fazly, Sven Wachsmuth: Learning Structured Appearance Models from Captioned Images of Cluttered Scenes
* Ondrej Chum, James Philbin, Josef Sivic, Michael Isard, Andrew Zisserman: Total Recall: Automatic Query Expansion with a Generative Feature Model for Object Retrieval
* Peter Hansen, Peter Corke, Wageeh Boles, Kostas Daniilidis: Scale-Invariant Features on the Sphere
* Tianli Yu, Ruei-Sung Lin, Boaz Super, Bei Tang: Efficient Message Representations for Belief Propagation
* Victor Lempitsky, Carsten Rother, Andrew Blake: LogCut - Efficient Graph Cut Optimization for Markov Random Fields
* Amir Tamrakar, Benjamin Kimia: No Grouping Left Behind: From Edges to Curve Fragments
* Liangliang Cao, Li Fei-Fei: Spatially coherent latent topic model for concurrent object segmentation and classification
* Stefan Roth, Michael J. Black: Steerable Random Fields
* Andrew Rabinovich, Andrea Vedaldi, Carolina Galleguillos, Eric Wiewiora, Serge Belongie: Objects in Context
* Yan Ke, Rahul Sukthankar, Martial Hebert: Event Detection in Crowded Videos
* Per-Erik Forssen, David Lowe: Shape Descriptors for Maximally Stable Extremal Regions
* Sanjiv Kumar, Henry Rowley: Classification of Weakly-Labeled Data with Partial Equivalence Relations
* Simon Polak, Amnon Shashua: Latent Model Clustering and Applications to Visual Recognition
* Simon Prince, James Elder: Probabilistic Linear Discriminant Analysis for Face Recognition
* Tom Yeh, John J. Lee, Trevor Darrell: Adaptive Vocabulary Forests for Dynamic Indexing and Category Learning
* Florian Schroff, Antonio Criminisi, Andrew Zisserman: Harvesting Image Databases from the Web

Gaussian CRF

2007-08-13T15:06:00.000-04:00

Found an interesting paper that might be relevant to my work also.

Learning Gaussian Conditional Random Fields for Low-level vision
M. Tappen, C. Liu, E Adelson, W Freeman
CVPR 2007

I can not afford to delve more in this paper right now, but I dont want to miss it either.

ICCV 2007

2007-07-25T23:14:00.001-04:00

Hurray!!

The vision lab at UMass has two papers accepted at ICCV: my paper and Gary's paper. I put a DRAFT of my paper on my webpage.

Since the list of accepted papers is posted on the ICCV website, soon their will be queries for specific papers on different search engines (I hate monopolies, so hate to say the G-word). I think it is good for the research if people put a draft of their accepted paper online so that anybody who is interested in reading a particular paper does not have to wait till October to read it.

I remember I found the title of a paper very intriguing at CVPR 2007 but was very annoyed by the unavailability of an online version and the unresponsive authors.

A comparison of feature detectors and descriptors

2007-07-12T15:57:00.000-04:00

I (and other people like me) have spent a lot of time looking for the comparisons of local feature detectors and descriptors for images/videos. I feel ashamed that it took me this long to find this awesome tutorial given by Tinne Tuytelaars at ECCV 2006. I feel that either the bloggers in the computer vision are not doing their job or the search engines are not bringing up the right references, assuming that I did try searching for the references intelligently enough.

Anyways, the link is
here.

I hope this post will save atleast a few research hours.

CVPR'07 papers

2007-07-10T22:31:00.000-04:00

I promised myself that I would not read any more papers... but as always this resolution did not last very long. Over the time, I read a couple of papers from CVPR'07 but, due to that stupid resolution I did not post any notes.. Infact I am losing track of them as well.

lemme first list the papers I am interested in.. I hope to post my notes soon..
* Eric Nowak and Frederic Jurie, Learning Visual Similarity Measures for Comparing Never Seen Objects
* Simon Winder and Matthew Brown, Learning Local Image Descriptors
* Herve Jegou, Hedi Harzallah, and Cordelia Schmid, A contextual dissimilarity measure for accurate and efficient image search
* Marshall Tappen, Ce Liu, William Freeman, and Edward Adelson, Learning Gaussian Conditional Random Fields for Low-Level Vision
* Jun Li and Pengwei Hao, Hierarchical Structuring of Data on Manifolds
* Zhuowen Tu, Learning Generative Models via Discriminative Approaches
* Helmut Grabner, Peter Roth, and Horst Bischof, Eigenboosting: Combining Discriminative and Generative Information
* Boris Epshtein and Shimon Ullman, Semantic Hierarchies for Recognizing Objects and Parts
* Yair Weiss and Bill Freeman, What makes a good model of natural images ?
* Jieping Ye, Zheng Zhao, and Huan Liu, Adaptive Distance Metric Learning for Clustering
* Dashan Gao and Nuno Vasconcelos, Discriminant Interest Points are Stable
* O. Tuzel, F. Porikli, and P. Meer, Human Detection via Classification on Riemannian Manifolds

CVPR'07 paper from Riya

2007-06-13T20:03:00.000-04:00

title: Contextual identity recognition in personal photo album

I was looking at the papers at the coming CVPR. Being interested in contextual search and face recognition, this paper got me excited. From the electronic copy I found out that the paper is from Riya Inc. I always wanted to know some details of what is going at Riya, and here it is..

It is a decent systems paper but some details are missing e.g., face features. Though the authors say that it is beyond the scope of the paper, they are probably the most critical of the features; and presumably the part that has the maximum potential for engineering :D

Some details:
* use face + clothing features within a time frame ("event").
* Overall modeled as a markov network with:
- nodes as detected faces.
- three kind of potential functions for - face similarity, uniqueness of individual within a photo, and clothing similarity.
* Inference done using loopy BP with parallel sum-product updates.
* clothing features:
- adaptive binning using k-means over RGB components
- Earthmover's distance
- texture features: Gabor filters.
- regression with cloth similarity score for appropriate weights for above-mentioned components.

Experiments: their own six data sets. I wonder if those are publicly available.

Overall, the paper was OK. I wish there were more details (as always).

Mean Shift

2007-02-28T12:46:00.000-05:00

Thanks to my lab presentation, I got a chance to learn about mean shift algorithm (there are so many concepts that one hears about and add to the TODO list but never does it). I eventually found THE reference for learning it (and related concepts).

http://www.wisdom.weizmann.ac.il/~deniss/vision_spring04/files/mean_shift/mean_shift.ppt

It does explains theory and applications of mean-shift algorithm (with lots and lots of animations).

Ensemble methods

2007-02-14T19:59:00.000-05:00

For a long time, I was thinking of reading details about ensemble methods. Thanks to the snowstorm, I finally got some time to find a good survey article by Dietterich on ensemble methods in machine learning:

http://citeseer.ist.psu.edu/dietterich00ensemble.html

hiatus

2006-10-17T09:16:00.000-04:00

came across a lot of interesting papers in the past few days. unfortunately I am working on getting some stuff done so could not post notes on them. But before I forget about them.
* A collapsed variational bayesian inference algorithm for LDA (Teh et al., NIPS 06)
* ??

Look, I already forgot the others.

BMVC' 06

2006-09-07T19:47:00.000-04:00

Talks :
-------

* First order geometric distance (the myth of the Sampsonus)
- Sampson error is not the first order approximation of the distance to a curve, instead it is a distance to the first order approximation of the curve.

* Brian Curless's invited talk (cool demos about story-boarding, cool camera for
multi-focus capture and multi-view interpolation
* Coupling face recognition and super-resolution
- this aspect is interesting to incorporate into face reco from video.

* Specificity as a graph based estimator of cross estimator and KLD
- implicitly does the kernel density estimation (author wont accept it though :D)
- model evaluation should be done in a different way than what is used in model
selection or optimization criterion.

* Extracting scale and illuminant invariant regions thru color
- ??? (not sure if I am convinced)

* deformation and viewpoint invariant color histogram
- da (f_x*g_y - f_y*g_x) is invariant, where a is the area of the region with constant color, f,g and some channel signal (R/G/B) and f_i is the i^th derivative.
- there are clearly specified limitations and assumptions

* patch based object recognition using discriminatively trained gaussian mixtures
- looks similar to transformed dirichlet processes (NIPS'05) with gaussian mixtures
instead of the dirichlet process

* Real time feature matching using adaptive and spatially distributed classification
trees
- looks good .. .not really clear about the entire process.

* Tied Factor analysis for face recognition across large pose changes
- might be interesting

* Hello! my name is buffy... automatci naming of characters in TV video (best industry prize)
- elaborate system with many components implemented in simplistic fashion
- enrich subtitle text by crawling the web and looking for scripts (often get info about the name of the speakers of particular text).
- use info from subtitles to make correspondences between text and timestamps in video
- appearance model for lips/mouth for speaker identification
- track different characters over time/frames
- also use some clothing information for better tracking

* Finding people in repeated shots of the same scene
- assumption: people are not changing clothes and hair style over this duration when repeated shots are taken.
- works when faces are not frontal in some of the images
- spring model for torso clothes + face + hair

* An algo for tuning an active appearance model to new data

* Feature detection and tracking with constrained local models (Best science prize)

Posters:
--------
* ofcourse mine was the best :D

* video super-resolution with scene specific priors

* incremental learning of locally orthoonal subspaces for set based object recognition
- unordered sets. effectively incremental LDA. would fall apart for large number of
classes. experiments with 100 classes.

* motion as shape : a novel method for the recognition and prediction of biological motion
- shape matching for motion trajectory. looks similar to matching flexible character models for handwriting recognition.

* semi-supervised learning of joint density model for human pose estimation
- to emphasize that unsupervised learning is useful (ain't there many papers doing this already :D )
- tracking experiments

* Preconditioning for temporal video superresolution
- interesting concept (to read ...).. should work just as fine for spatial superresolution
- circulant matrices

* learning distances for arbitrary visual features
- couldnt see this poster, may be interesting

* efficient clustering and matching for object classification (Mikolajczyk et al)
- must look

* Compact object descriptors from local invariant histograms
- eight parameters of weibull distribution are learnt. (details ??)

* Automatic video segmentation using spatio-temporal t-junctions (best poster award)

* Field of experts for image-based rendering (Fitzgibbon)
- interesting energy formulation as a sum of energy based on natural image stats and normal FoE formulation.
- final image is rendered by picking clusters from one of the local minima of two energy functionals.
- must read the details

* segmentation based multi-cue integration for object detection
- ??

Non-Negative Matrix Factorization

2006-08-11T13:08:00.000-04:00

People who know about PCA (which consists of almost everyone), especially those who are new to this field, should also consider learning about non-negative matrix factorization. This often helps in better understanding of data and problem domain.

Relevant links

http://en.wikipedia.org/wiki/Non-negative_matrix_factorization

CVPR'06 notes contd.

2006-06-25T13:22:00.000-04:00

- A new method of probability density estimationwith application to mutual information based image registration
nice results. ??

- The bottleneck geodesic: computing pixel affinity.
interesting. MUST read

- graph laplacians kernels for object classification from a single example
pseudo-inverse of normalized graph laplacian is the reproducing kernel of H(G)

- discriminative object class models of appearance and shape by correlograms.
what are correlograms useful for?

- learning object shape: from drawings to images
(from NIPS workshop) might be interesting

- region based image annotation using asymmetrical SVM based multiple instance learning

- Satellite features for the classification for visually similar classes (epshtein)
local coordinate system for registration.
identification of 5 actresses
compare with hyperfeatures.

- putting objects in perspective (derek hoeim) best paper award
nice..
detecting pedestrians and cars using surface detection simultaneously.
must read the details.

- context and hierarchy in a probabilistic image model (geman)

CVPR'06 notes contd.

2006-06-25T12:50:00.000-04:00

- Noise estimation from a single image
missed it... must look
- extracting subimages of an unknown category from a set of images
- augmenting shape with appearance in vehicle category recognition.
UIUC cars dataset, another one at brown.edu
- composite templates for cloth modeling and sketching (Song Chun Zhu)
stochastic context sensitive grammar.
modeling caricatures

- weekly supervised top-down image segmentation

- AnnoSearch: image auto-annotation by search (MSR Asia)
enriching annotation. input : image+simple annotation. compute feature encoding for the input image and retrieve similar images (alongwith annotations) from the data set and choose the top annotations to enrich the annotation of the query image.
naive approach. feasible only if you have huge annotated data.

- Hidden CRF for gesture recognition
linear chain CRF with another node (sub-category) connected with the hidden nodes. This node is observed. what about the semi-supervised case when this node is not observed for many training examples. (should look at their previous paper for details of this model)

- Using multiple segmentations to discover objects an their extent in image collections (Alyosha)
connect this to zoubin's paper.

CVPR'06 notes

2006-06-25T12:17:00.000-04:00

To read:
- principled hybrids of generative and discriminative models.
- accelerated kernel feature analysis

- escaping local minima through hierarchical model selection : autamatic object discovery, segmentation and tracking in video (Nebojsa)
kinda learning weights for different graph structures. (read details)
- spectral methods for automatic multiscale data clustering (zoubin)
only need to do one SVD and then take powers of the eigs to obtain multiple clustering. interesting (theoretical) idea, need to figure out where we can use it.
- training deformable models for localization (Deva Ramanan)
may be useful if I ever decide to go into tracking (impressive talk at the workshop on the last day)
- Diffusion distance for histogram comparison.
like earth-mover distance but fast. useful for SIFT matching.
- Transformation invariant component analysis for binary images
binary PCA (L. Saul et al. AISTATS 03)

- Picture Collage
- The design of high-level features for photo quality assessment (these two papers from MSR Asia : interesting applications)
-making a long video short: dynamic video synopsis

CVPR'06 workshops notes

2006-06-25T11:59:00.000-04:00

* Shree Nayar's invited talk in MMBIA workshop was really interesting. mostly graphics but may be useful for vision if the related ideas are developed for single image / video.
ideas : can separate global illumination and direct illumination components by having high frequency(spatial) light and using a checkerboard.
- see his SIGGRAPH'06 papers for more details.

* Beyond Patches
- Discriminative patch selection using combinatorial anad statistical models for patch based object recognition (missed the talk, but MUST see as related to what I am doing)
- Models for patch based image restoration
deblurring/blind image deconvolution. alternating NBP and BP for multilayer MRF.
- Evaluation of intensity and color corner detectors for affine invarient salient regions
interesting stuff: may use these color features for hyperfeatures.
- Using spatio-temporal patches for simultaneous estimation of edge strength, orientation and motion.
nice results. not yet convinced. motion detection in faces ?
- Morse functions for activity classification using spatiotemporal volumes.
not impressed. Iwasawa decomposition.
- Strangeness based feature selection for part based recognition
- dont remember.. must look into it

* RANSAC
- interesting stuff.... need to read more about it.

CVPR 06

2006-06-23T13:39:00.000-04:00

a reminder for myself: I should put my notes here before they are destroyed.

TODO list

2006-06-04T18:01:00.000-04:00

* Chapter 1-3 of M. Wainwright's thesis
* complete gamogistic analysis
* details of MRF, CRF
* more
* even more...

Interesting faces in News (CVPR 06)

2006-03-27T20:37:00.000-05:00

despite my strong resolutions of blogging, it seems I am drifting away pretty easily.

Anyways, today I read a paper from CVPR 06, which is quite relevant to my research- "Interesting faces in news". Honestly, I dont feel very great after reading this paper. We were planning to do exactly the same experiments but did not do them coz we could not find mathematical justification for it. And this paper talks about the experiments but does not assuage the inquisitive minds.

The approach is pretty much as following:

For a query name
* choose images that have this name (and already known other names) in the caption
* detect faces in these images
* make a graph with nodes as detected faces and edge weights as the average "SIFT" distance between the two face images
* use a greedy approximate algorithm to get the densest component of the graph (this is the output)

The thing I learnt from this paper:
* The reference to the greedy approximation algorithm
http://www.cs.princeton.edu/~moses/papers/greedy-dense.ps

Is CVPR about getting results or creating interesting perspectives/algorithms?

Alas!

Probability Theory

2006-03-14T16:22:00.000-05:00

Just like my previous post on my notes on measure theory, this one is to announce the beginning of a document on probability theory. Today I learned about some convergence theorems and inequalities. I knew about some of this stuff before but the rigorous treatment is interesting and helpful in clearing any doubts (hopefully).

the link
http://vis-www.cs.umass.edu/~vidit/notes/notes_probtheory.pdf

Measure Theory

2006-03-11T14:41:00.000-05:00

I started looking for some stuff in "Fundamentals of Statistical Exponential Families" and found that I need a refresher for some of the definitions from Measure Theory, Oh boy, I hate my handwriting. This happens to me recurringly so I decided to LaTeX my notes and why not make it public to help poor souls like myself.

http://vis-www.cs.umass.edu/~vidit/notes/notes_measures.pdf

Also the link is added to my "ambitious" plan of collecting stuff on "mathematical foundations of machine learning" webpage
http://vis-www.cs.umass.edu/~vidit/mfml.html

Mathematical Foundations of Machine Learning

2006-02-26T13:00:00.000-05:00

I have started a new page on the mathematical concepts that I feel everyone who is interested in machine learning should know. I will try to recollect all I have learnt in the past 2 years in top-down fashion. I personally feel that it would have been much better if somebody could have just told me about these things before so that I could have an unbiased understanding of these things. Lemme see if I can help someone in doing this. Beware I need a big a$$ acknowledgement. just kidding.

Oh I forgot the link
http://vis-www.cs.umass.edu/~vidit/mfml.html

All you need to know about logistic regression

2006-02-24T17:36:00.000-05:00

I am trying to put together all the knowledge that I have gained about logistic regression. Just uploaded a draft at
http://vis-www.cs.umass.edu/~vidit/notes/notes_logistic.pdf
I plan to update it as soon as I learn or observe something about it.

This is an attempt to understand what is actually going on while conditional training a generative model. As Tom minka says, "Discriminative models, not discriminative training", this document basically discusses specific case of conditional training of mixture models. Also, it is the qwest for the answer to the $100 mn question : "Why am I using this model and why should I not use the other one?"

Vision Conferences

2006-02-15T18:25:00.000-05:00

As Ashutosh suggested, here are the deadlines for vision/image processing conferences. Specific dates are not yet out

* ICASSP Sep/Oct
* CVPR Nov
* ICIP Jan
* ICPR Jan
* ICCV Mar

Exams over

2005-12-21T20:07:00.000-05:00

I am a little excited about the exams being over. There was too much multitasking going on in my mind for the last couple of days. Atleast some of it is alleviated. The next aim is to continue the experiments on hyperfeatures and review (with notes) the annotation literature.

Before I forget, have to go through some stuff on information theory that I have been postponing for quite a while.

... and miles and miles to go before I sleep

Upcoming Deadlines

2005-12-17T00:20:00.000-05:00

ICML 06 : Jan 30
AAAI 06 : Feb 16
UAI 06 : March 9
NIPS 06 : June

NIPS day 3

2005-12-09T18:05:00.000-05:00

talk : Diffusion methods, spectral clustering
-------
probabilistic interpretation of spectral methods.

talk : distance metric learning fo large margin NN (entertaining graphics in presentation)
----------
- loss function (L) = push force + pull force (not convex)
- hinge loss on push force
- However, M = L^T*L is convex, use semidefinite programming to minimize with M
- classification - (a) Mahalanobis kNN (b) minimum loss classification
- impostor based loss function (Chopra, CVPR05)

Talk: Estimation of intrinsic dimensionality using high resolution vector quantization
---------------
Wasserstein distance is the continous generalization of Earthmover's distance.

Posters
-------
* Non gaussian Component Analysis
- reconstruct the non gaussian component of given high dimensional data.
- interesting paper (to read)

* Generalization to unseen cases
- empirical error can be used to bound the off-training-set error.
- neat and simple idea

* Metric learning by collapsing cases
- learning Mahalanobis metric
- minimize KL-divergence with the ideal situation in which the within class
distance is zero and interclass dstance as infinity.
- can use kernel to map conentric distributions into a space where the mapping
results in distinct "collapsed" centers.

* A probabilistic approach for optimizing spectral clustering
- known number of clusters
- proabibility of ith cluster, p(z_i) = # edges within cluster i / sum of degrees for the nodes in the ith cluster

* Inference with minimal communication : a decision theoretic variational approach
- belief propagation with a constraint that any node can only pass a bit as the message
instead of a real valued message in the original BP.
- application to sensor network

* From Lasso regression to Feature vector machine
- modeling Lasso into a SVM framework and thus derive the convex optimization
framework.
- observation : SVm wants every data point to be outside the margin (in data space)
lasso wants every data point to be inside the margin (in feature space)
coz we assign non-zero weights to informative features only
- once we have formulated a svm type formulation, use kernel methods to add flexibility on decision boundaries.
- To read .. looks like an interesting idea

* Affine structure from sound (Sebastian Thrun)
- localizing a set of microphones with sources of sound with unknown locations and associated time
- based on Tomasi's work on Structure from motion.
- uses simple tricks from geometry and algebra (SVD)

NIPS day 2 posters

2005-12-07T11:51:00.000-05:00

* Correcting sample selection bias in maxent density estimation
addresses the issue of estimating the unbiased statistics when the sample selection bias is known. This is not directly related to what we are trying to do as trasfer of parameters from training to testing when we know the prior distribution for test cases.
useful reference : Bianca Zadrozny (ICML04)

* Two view learning: SVM-2K, theory and practice
combines two svm that work on different features. does a joint optimization for the selection of support vectors in the two classifiers

* Mixture Modeling by Affinity Propagation (Brenden Frey)
soft clustering method using pair-wise affinities.
might be interesting. good presentation. need to read paper.

* Tensor Subspace Analysis (He, Niyogi)
Instead of considering image patches as a vector of pixel values, treat it as a tensor maintaining the 2D neighborhood. treatment very similar to PCA.

* Spectral bounds for sparse PCA: Exact and Greedy algorithms (Moghaddam)
To read. an interesting and thorough approach for discrete version of spectral methods.

* Multiple instance boosting for object detection (Viola)
- for training we need labels for a bag of data points with the constraint that a positive label implies atleast one positive example in the collection.
- Noisy-OR for MIL
- Boosting for logistic bag regression
- apply anyboost to reweight data

NIPS : day 2

2005-12-07T11:48:00.000-05:00

Talk : Urs Holzle (Google)
-------
* stuff about Google, results from NIST competition which Google won.
* GFS

Talk : Pradeep (Preconditioners)
--------------------------------
approximate inference techniques for graphical models using preconditioners
upper and lower bounds on graphical model energy(?)

* Vaidya (1990) seminal work on using preconditioners

* alpha-expansion algorithm :
Boykov, Y., Veksler, O., Zabih, R.. Fast Approximate Energy Minimization via Graph Cuts.
In IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI). vol. 23,
no.11, pp.1222-1239, 2001

Talk : Jaety Edwards
---------
Searching for character models
looks similar to Viterbi algorithm ??? (to read)

Talk : Computational model of eye movements during object class detection
-------
may not be useful... need a brief look at it

Talk : Top-down control of visual attention.
------
HMM-like model... brief look

NIPS : day 1 posters

2005-12-06T13:40:00.000-05:00

Interesting posters:
* Erik Sudderth et al., Describing visual scenes using Transformed Drichlet Processes
* A. Saxena et al. Leaning depth from single monocular images
* Y. Bengio et al. Non-local Manifold parzen windows
* M. Narasimhan, Q-clustering

NIPS 2005 : Day 1

2005-12-06T12:12:00.000-05:00

Tutorial: Lawrence Saul
----------------------
Spectral Methods : ISOMAP, LLE, Hessian LLE, maximum variance unfolding
Stokes' Law : squared gradient = function * laplacian
To see:
MDS
Hessian LLE (Donoho)

??
* look for the surfaces where the tangent space is more well-behaved as compared to the
original surface.
* how to verify the presence/absence of the holes in the data.
-- no principled approach, hit and trial(Baback Moghaddam)
-- references from CVPR/face recognition for use of spectral methods

Tutorial : Stuart Russell, Brian Milch
--------------------------------------
Good introduction
BLOG - bayesian logic model (similar to the talk at Umass)

Tutorial : Michael Jordan
----------------------
Measure Theory
Probability theory
Dirichlet Process (Chinese restaurant problem)
Hierarchical Dirichlet Process (Chinese restaurant franchise)
Discussion about Bayesian vs frequentist approaches (pros and cons)
-- Diaconis, Freedman (96??) - Incorrect priors lead to wrong models even with the
infinite data
Interesting concluding slides on philosophical discussions

Linear Regression Contd....

2005-06-24T13:29:00.000-04:00

In linear regression, we often want to add a penalty term to the sum of squared error term to capture the semantics of the problem domain.

The two most common ways of doing so are:

Ridge Regression : penalty term --- \SUM_i | \beta_i |^2
Lasso : penalty term --- \SUM_i | \beta_i |

In lasso you get exactly zero values for the betas which correspond to the least relevant component

LARS (Least Angle Regression)

To obtain the regression solution, the usual approach is to estimate \hat{\mu} = X \hat{\beta}. This is often done through an iterative update for \hat{\beta}. The two most common methods that were used in the past were Forward Selection and Stagewise. Forward Selection adopts a greedy approach and picks the most correlated covariate and updates \hat{\mu} with the projection of response vector Y onto that covariate at every iteration. Stagewise, on the other hand, updates \hat{\mu} by \epsilon multiple of the projection.
Forward Selection has the problems related to the greedy approach whereas Stagewise can be very slow.

Least Angle Regression borrows ideas from these two approaches. We start with all coefficients equal to zero, and find the most correlated covariate with the response. We update \hat{\mu} with the maximum possible multiple of this covariate until some other covariate has equal/more correlation with the residual response. From this point, we move along the direction that is equiangular between the two covariates in the similar fashion. At every step we are adding a new covariate into consideration, hence the total number of steps required is equal to the number of participating variables.

LARS can be easily modified to solve ridge regression and lasso. It provides an efficient way of solving regression.

Linear Regression

2005-06-14T09:37:00.000-04:00

Linear : The regression function E(Y|X) i slinear in the inputs X_1, X_2, ...,X_p

Linear Regression methods:

OLS (Ordinary Least Squares)

minimize the residual sum of squares, RSS
RSS(b) = SUM_i=1^N (y_i - f(x_i))^2
= SUM_i=1^N (y_i - b_0 - SUM_j=1^p x_ij b_j )^2

OLS is reasonable if (x_i,y_i) represent independent random draws from their population or if the y_i are conditionally independent given inputs x_i

Solution : RSS(b) = (y-Xb)^T (y-Xb)
min RSS(b) => b = (X^T X)^-1 X^T y

Gauss Markov Theorem

The least squares estimate of the parameters b have the smallest variance among all linear unbiased extimators.

Any method that shrinks or sets to zero some of the least square coefficients may result in a biased estimate.

Statistical Learning Theory

2005-05-28T00:16:00.000-04:00

I was trying to do the analysis of some variant of SVM. In the course, I felt the need of the statistical learning tools.

This made me realize that one should be familiar with the Probability and statistics to reason in detail about any form of learning (to speak in a trivial fashion). However, if you really want to understand machine learning, you should read the book on Statistical Learning Theory by Vapnik. I feel this book as the perfect balance between the concise formulation and descriptive discussions.

There are other useful references one should refer

Statistical Inference - Casella & Berger
A Probabilistic Theory of Pattern Recognition- L. Devroye, L. Gyorfi, and G. Lugosi

Transformation-Invariant Embedding

2005-04-03T17:33:00.000-04:00

I came across the paper by Ghodsi et.al. on Transformation-Invariant Embedding for Image Analysis.

It discusses the alternatives for the Euclidean distances in Manifold Learning to encompass the perceptual understanding/differences.

Will discuss about it more later....

Differential Geometry Basics

2005-03-22T00:41:00.000-05:00

Just to remind some basics about manifolds:

Homeomorphism : bijection (theta) between topological spaces for which (theta) and (theta^-1) are continuous.
Charts : Local homeomorphisms on Manifolds.
Diffeomorphism : bijection f : M -> N, such that f \in C^\infty (M,N) and (f^-1) \in C^\infty (N,M)
Tangent Space : composed of directional derivative operator operating on C^\infty(M,R) functions. Every vector in the Tangent Space is an operator.
Derivative of a function : At point p, f_(*p) : T_p(M) -> T_f(p)(N)

forall : N -> R, f_*p v = v(r o f)

Locality Preserving Projections

2005-03-16T23:21:00.000-05:00

Today I read the NIPS paper by He and Niyogi. Sounds like a nice result. Still need to be comfortable with the intuition behind the beautiful mathematical formulation.

nothing much...

2005-03-15T13:28:00.000-05:00

lets start bloggin.... must keep rollin' am catching too much moss