Outobox - Machine Learning Reading Group


This is an interdisciplinary group of faculty (Arindam Banerjee, Dan Boley, Maria Gini, Paul Schrater, and William Schuler) and students who meet every week to discuss research ideas and to share relevant literature. The group is called Outobox, because we like to think ...out of the box. There is a mailing list associated with this group. If you would like to be added, please go to sign-up page.

There is a brand new discussion forum associated with this group. Please contribute to the discussion.


Fall 2009

This semester the meetings will be on Tuesday from 4:00 to 5:30 in EE/CS 6-212.
  1. December 8 -- Tim Miller will be speaking on his research as practice for a faculty job talk.
    The topic will mainly be his thesis research on syntactic processing of speech repair with time-series models (Hierarchical HMMs), with additional focus on how he sees this research expanding in the next 1-3 years into unsupervised learning of syntactic structure, integration of syntax with other linguistic information, and explicit psycholinguistic modeling.
  2. December 2 -- Tom Vacek will talk on "Using Wikipedia for Topic Extraction"
  3. November 23 -- David Olsen will talk on Analyzing the Effects of Higher Dimensionality on Various Clustering Methods.
  4. November 17 -- Rudy Agovic will present the paper Neil Lawrence, Probabilistic non-linear principal component analysis with Gaussian process latent variable models, Journal of Machine Learning, 2005.
    Link: http://jmlr.csail.mit.edu/papers/volume6/lawrence05a/lawrence05a.pdf
  5. November 10 -- Urvashi Mishra on "Building a Tool to Track the Geographic Spread of the Flu Virus"
    We discuss an interactive tool to track the geographic spread of the flu virus, including some of the issues in designing and building the tool, and some snapshots of a prototype.
  6. November 3 -- Elizabeth Jensen on "Long-Term Multi-Robot Patrol of an Open PolyLine".
    Abstract: Multi-robot patrol is a growing field in which a team of robots works to optimize the frequency with which they pass through certain points. Applications include garbage collection, street cleaning, and surveillance. Previous work has been focused on the uniformity of the patrolling algorithm. We present an algorithm that focuses on long-term maintainability of a multi-robot patrol system. The objective is for a team of robots to patrol an open polyline, such as a fence or border. Each robot is assigned to either patrol a segment of the line or queue in the home base as a reserve robot. When a patrolling robot runs low on power, it removes itself from the line and returns to the charging station, while a reserve robot is deployed and the other patrolling robots reorganize to cover the open segment and make room for the new robot. Once a robot is recharged, it is added to the queue of reserve robots and waits until it is needed. The robots run autonomously for the most part, with communication limited to interaction between an individual robot and the home base to coordinate organization changes and record state information. The system can maintain coverage of the full line so long as there are enough robots left active to cover each segment. In this paper we show, through both physical experiments and simulations, that the algorithm maintains coverage of an area over a long period of time due to the ability to recall, recharge and reorganize the structure of the robot team.
  7. October 28 -- no talk
  8. October 14 -- no talk, there is a robotics talk at 4:00 in Walter 401
  9. October 7 -- baylor wetzel will talk on "The Whole Town Is Talking: Using Abstraction, Intention and Composition to Quickly and Easily Create Large Numbers of Unique, Reactive Conversational Agents"
    Alelo makes video games that teach people culture. Like many companies, they have tools to make intelligent, believable agents but not intelligent, believable cities. Because it takes one-to-three weeks to make a single conversational agent, time, money and manpower make it essentially impossible for most companies to make massively multi-agent systems. In this talk, i'll discuss a set of techniques used to greatly scale the AI authoring process. We'll cover the work on conversational agents i did for Alelo's Virtual Role Players system and the ongoing work i'm doing on fashion for Shikigami's PaperDoll Fashion Reaction game. The talk will focus on using composition and model-by-exception design patterns to build individuals from an arbitrary set of cultures, where a culture is modeled as a sparse set of response types. Note that this talk will focus on agent intent and will only briefly discuss behavioral realization.

    The talk is based on this paper (Game Programming Gems 8, 2010).

    Background: Sometimes, quantity is quality. Many times a game maker has dreamed of making a virtual world filled with thousands of intelligent agents. If done correctly, these computer-controlled characters act just like people, allowing the player to experience phenomena they could not experience in smaller scale systems. Multi-agent and massively multi-agent systems of behaviorally plausible agents provide more than entertainment, they allow researchers to study phenomena such as influence propagation, tipping points, steering, contextual decision making, group decision making, individual differences, culture, leader theory and much more. Groups such as the University of Pennsylvania, the Army Corps of Engineers and the Navy have used such systems to study the behavior of people fleeing a fire, soccer hooligans, terrorist cells, riots, inner city gangs and the reaction of Afghan villagers to U.S.-led reconstruction activities. Despite the value inherent in massively multi-agent systems, few such systems exist. Those that do typically suffer numerous, significant undesired compromises - agents lack individuality or personality, have limited knowledge, have no persistent memory, are not sufficiently context aware, etc. The reason for this is cost. At Alelo, a company that makes cultural training games using conversational agents, it took two people (a designer and programmer) a combined three weeks to build a single conversational agent. This was only enough time to produce an agent capable of performing a single "conversational task". Populating a town with hundreds of unique, functional characters requires too much time, money and manpower.

  10. September 29 -- no meeting
  11. September 22 -- Will Groves will talk on Price Prediction Methods for Real and Simulated Goods Markets".
    Price prediction is a useful activity in many domains from the physical realm of supply chain management to the very abstract domain of financial products. Predictions, if of sufficient accuracy, can be used to profitably decide market timing, can indicate that a speculative opportunity exists, and can be used to decide how best to hedge against risk. There are many prediction techniques described in the literature, and each technique has different requirements for data format and quality. This work addresses the question of how much data preparation is required to run effectively given real data. We use data from both real commodity markets and from the Trading Agent Competition in Supply Chain Management (TAC SCM). Performance results from various popular off-line (data mining) and on-line (time series) prediction techniques are presented.
  12. September 15 -- Steve Damer will talk on "Cooperating with an opponent in a persistant changing environment"
    I will discuss my perspective on the problems of acting in a persistent changing multi-agent environment, a model to solve some of those problems, and a learning system to enable intelligent use of that model. In a multi-agent environment, it is important to consider the effect of your actions on your opponent's states, because your opponent's actions can affect your payoffs. However, you are unable to observe your opponent's state - you are not even given a model of your opponent. Even if you were given a model of your opponent, you would likely lack the computational resources to use it. I propose attitude, belief, and method for action selection as a model of behavior which allows for cooperation, while remaining simple enough to reasonably model. I describe how a pre-regularized particle filter can be employed to learn attitude, belief, and method. I show some examples of how the environment affects the rate of learning.

Spring 2009

This semester the meetings will be on Friday from 2:00 to 3:30 in EE/CS 5-212.
  1. April 24 -- Steve Damer will talk on "Extending Regularized Particle Filters with Discrete Parameters". He will talk about extending a standard regularized particle filter which predicts attitude and belief to also predict method of picking a Nash equilibrium
  2. April 17 -- David Olsen will talk on the paper "When Is 'Nearest Neighbor' Meaningful?" by Kevin Beyer, Jonathan Goldstein, Raghu Ramakrishnan, and Uri Shaft, Lecture Notes in Computer Science, Database Theory -- ICDT'99 Volume 1540 (1999), pages 217-235
    http://www.springerlink.com/content/04p94cqnbge862kh/ (accessible only from U of M hosts)
  3. April 10 -- Ham Ching Lam on on data mining on avian flu data.
  4. April 3 - Dimitri Jevremovic will talk on "MicroRNA prediction with a novel ranking algorithm based on random walks" based on paper located at: http://bioinformatics.oxfordjournals.org/cgi/content/full/24/13/i50
    I will focus on the major idea of the paper, which is how to use semi-supervised learning to classify (predict) large data sets based on very few known samples.
  5. March 27 -- Will Groves and Mohamed Elidrisi will talk on their work on predicting prices in the trading agent competition for supply chain management and in the commodity market.
  6. March 6 -- Steve Damer will talk about particle filters.
  7. February 27 -- John Josephs will talk about his work on "Discovering generalized concepts from documents using the Wikipedia category hierarchy"
  8. February 20 -- Steve Damer will talk about his work on cooperation of agents.
    We describe a simple environment to study cooperation between two agents. The environment consists of randomly generated normal form games with uniformly distributed payoffs. Agents play multiple games against each other, each game drawn independently from the random distribution. An agent identifies cooperative moves by assigning an attitude to its opponent and itself. The attitude determines how much a player values its opponent payoff, i.e how much the player is willing to deviate from strictly self-interested behavior. To cooperate, an agent estimates the attitude of its opponent by observing its moves and reciprocates by setting its own attitude accordingly. We show how this can be done using a particle filter, even when the opponent is changing its attitude."
    More details in the paper
    Steve Damer and Maria Gini, Achieving Cooperation in a Minimally Constrained Environment, Twenty-Third Conference on Artificial Intelligence (AAAI 2008), Chicago, July 2008, pp 57-62.
  9. February 13 -- Gyan Ranjan will talk on random walks on graphs.
    "Random walks on graphs have found useful applications in multiple domains, ranging from collaborative recommendation tasks to opportunistic routing in ad-hoc networks. One can find a vast literature in the field of electrical networks which relates random walks to the electrical properties of an undirected graph. For example, the expected one way transit and commute times of a random walk between a pair of nodes in the graph are closely related to the pairwise effective resistance value between the two nodes. Such measures can be computed using the Moore Penrose pseudo inverse of the Laplacian of the graph which has many interesting properties. This talk will deal mainly with a PCA kind of analysis on an undirected graph using the expected commute time measures. We will also see how aggregates defined over this quantity can be used to infer structural properties of a graph. In essence, we will cover previously known results which formed a basis for Dr. Boley's talk in the first session. "
    Reference paper by M. Saerens, F, Fouss, L. Yen, and P. Dupont, "The Principal Component Analysis of a Graph and its Relationship to Spectral Clustering.".
    This paper is closely related to the paper F. Fouss, A. Pirotte, J-M. Renders, and M. Saerens, "Random-Walk Computation of Similaraties between Nodes of a Graph with Application to Collaborative Recommendation". IEEE Trans on Knowledge and Data Engineering, Vol 19, N. 3. March 2007, pp 355-369.
  10. February 6 -- Tim Woods will give the talk that was scheduled for Januray 30.
  11. January 30 -- CANCELED DUE TO BUILDING CLOSING
    Tim Woods will talk about his work on security in machine learning, specifically in spam filtering.
    For more information, read the paper he wrote with his project team for a class last semester: Security in Machine Learning: measuring the relative sensitivity of classifiers to adversary-selected training data by T. Woods, M. Evans, D. Rust, and B. Podoll.
  12. January 23 -- Dan Boley will talk about "Random walk affinities for fully connected directed graphs."

    This talk is based on the paper "Random-Walk Computation of Similarities between Nodes of a Graph with Application to Collaborative Recommendation" by Francois Fouss and Jean-Michel Renders. That paper shows that the pseudo-inverse of the Laplacian of an undirected graph (a) gives the expected round-trip commute time between any pair of vertices, (b) matches the effective resistance between two vertices when the network is treated as an electrical network, (c) shows how this commute time is an euclidean metric for a particular embedding of the vertices in Euclidean space. As a result, the paper claims experimentally that this leads to more accurate affinity measures in collaborative filtering. In my talk, I summarize some of the theoretical results and show how some of these may be extended to fully connected directed graphs.

    The paper is available at http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=4072747&isnumber=4072743

    @article{Fouss07,
      author = {Francois Fouss and Jean-Michel Renders},
      title = {Random-Walk Computation of Similarities between Nodes of a Graph
               with Application to Collaborative Recommendation},
      journal = {IEEE Trans. on Knowl. and Data Eng.},
      volume = {19},
      number = {3},
      year = {2007},
      issn = {1041-4347},
      pages = {355--369}
    }
    

Fall 2008

This semester the meetings will be on Friday from 11:00 to 12:30 in EE/CS 6-212.

Spring 2008

This semester the meetings will be on Wednesday from 3:30 to 5:00 in EE/CS 6-212.

Fall 2007

This semester the meetings will be on Wednesday from 3:00 to 5:00 in EE/CS 5-212.

Spring 2007

This semester the meetings will be on Thursday from 2:15 to 4:00 in Walter 405. First meeting on January 25 will be in room 404.

Fall 2006

This semester the meetings will be on Wednesday from 3:15 to 5:00 in Walter 405. First meeting on September 13.

Past year meetings

  • June 29 -- Tim Miller has presented a demo of the new speech recognition system developed in the NLP lab and talk about its architecture.
  • June 15 -- Arindam presented his ICML paper.
  • June 1 -- Wolf Ketter did a dry run of his job interview talk.
  • 27 April 2006 - Dongwei Cao talked about his work.
    On Approximate Support Vector Machine and its Generalization Performance

    We study the algorithms that speed up the training process of support vector machines by using an approximate solution. We focus on algorithms that reduce the size of the optimization problem by extracting from the original training data set a small number of representatives and using these representatives to train an approximate SVM. Using the approach of algorithmic stability, we prove a PAC-style generalization bound for the approximate SVM, which includes as a special case the generalization bound for the exact SVM, which is trained using the original training data set.

  • 20 April 2006 - no meeting
  • 13 April 2006 - no meeting
  • 6 April 2006 - Evangelos Theodorou talked about the newer inverse reinforcement learning by Andrew Ng. Paper.
  • 30 March 2006 - Evangelos Theodorou talked about inverse reinforcement learning. Paper
  • 23 March 2006 - Tim Miller talked about hierarchical Hidden Markov Models (HHMMs), their relation to Dynamic Bayes Nets (DBNs), and how they are used in the NLP lab. Paper - Slides
  • 16 March 2006 - Spring break - no meeting
  • 9 March 2006 - William Schuler talked about DBNs and the speech recognition odyssey of the NLP lab.
  • 2 March 2006 - Steve Damer will be talking about his recent work.
  • 23 February 2006 - .. and now we move to MRF. The paper is http://www.cs.cornell.edu/~rdz/Papers/BVZ-cvpr98.pdf. Paul will lead the discussion.
  • 16 February 2006 - Now that we have learned everything about CRFs in theory, we will look at an application to computer vision. The paper we will look at is Using the Forest to See the Trees: A Graphical Model Relating Features, Objects, and Scenes, by Murphy, Torralba, and Freeman. Nate Bird will be presenting the paper.
  • 9 February 2006 - We will move forward to Conditional Random Fields (CRFs) by reading the paper Conditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data by Lafferty, McCallum, and Pereira. Other resources are two tutorials (by (1) Wallach and (2) Sutton and McCallum).
  • 2 February 2006 - We will read and discuss a paper on Maximum Entropy models, and if time permits we will study their relevance to predicting how many more weeks of winter we will have. There is a web tutorial on Adam Berger's webpage, as well as a paper on applying maxent models to NLP which is similar to the tutorial through the first 4 sections.
  • 26 January 2006 - For our first meeting this semester, Dongwei Cao will be talking about the work he will be presenting at the SIAM data mining conference, "On Approximate Solutions to Support Vector Machines."
  • 8 December 2005 - This week we will be looking at the paper "Factor Graphs and the Sum-Product Algorithm" by Kschischang, Frey, and Loeliger. Link
  • 1 December 2005 - We decided to do a review on Hidden Markov Models (HMMs). We covered them last year, but this will help get new people up to speed and the old people to clear out the cognitive cobwebs. This is all in anticipation of learning more about undirected graphical models like Markov Random Fields, and then possibly moving to Conditional Random Fields, and after that we should know everything about machine learning. Link to Rabiner paper
  • 24 November 2005 - No meeting - Thanksgiving holiday. Remember to give thanks for the outobox webmaster.
  • 17 November 2005 - Continuation of last week's topic of graphical models.
  • 10 November 2005 - Towards the end of last weeks session we started talking about the application of Markov Random Fields to Miguel's problem of detecting proteins from electron microscopy images. We had so much fun we decided to look a little bit at MRFs and graphical models in general. So, for next week, we will be reading a tutorial by Kevin Murphy which you can find at this location. There is also a tech report by Wainwright and Jordan at this location (postscript format) which is quite long, but fear not! The intro is a nice overview of applications of graphical models and is quite readable I'm told.
  • 3 November 2005 - Miguel, a visiting graduate student from Spain, will talk about his work done there which he is continuing here.
  • 27 October 2005 - We will again not be having a meeting, in observation of Theodore Roosevelt's birthday. If you would like something to mull over the two-week break, think about how you would build an agent to wash dishes by hand. In order to get a feel for the scope of the problem, you are welcome to visit my apartment and try the task out for yourself.
  • 20 October 2005 - We looked at a paper by Friedman, Hastie, and Tibshirani - "Additive Logistic Regression: A statistical view of boosting." Citeseer link is here. The diligent reader may also be interested in this paper "Why the logistic function? A tutorial discussion on probabilities and neural networks" by Michael I. Jordan.
  • 13 October 2005 - No meeting because of Navy Day Holiday
  • 6 October 2005 - We read an introduction paper on boosting, which helped to motivate and clarify the paper we read last week. Here is a link to a pdf version of the paper "A short introduction to boosting," and here is a link to the original adaboost paper, "A Decision-Theoretic Generalization of on-Line Learning and an Application to Boosting."
  • 29 September 2005 - Discussed the paper "Logistic Regression, Adaboost, and Bregman Distances" by Collins et al. (2001). Citeseer link
  • 22 September 2005 - Harini Veeraraghavan talked about her research regarding event detection in videos of automobile traffic.
  • 15 September 2005 - New faculty member Arindam Banerjee talked about his research on Bregman Divergences.
    Slides

  • 2004-2005

    Some people have asked for a list of papers the group looked at last year. If anyone has any additions let me know.