Skip to main content
    Israel [change]    Terms of use
    Home    Products    Services & solutions    Support & downloads    My account    
IBM Research

Machine Learning Seminar 2005

IBM Haifa Labs

Invitation Program Registration Abstracts

June 05, 2005
Organized by IBM Research Lab in Haifa, Israel

Some Disease Gene Hunting Algorithms

Prof. Dan Geiger, Technion

I will describe two methods for locating disease genes: one using pedigree information, called genetic linkage analysis, and the other using random affected and healthy individuals, called association analysis. Both methods use Bayesian networks in their internal representation and require learning structure and/or parameters. This talk is based on several works with G. Greenspan, M. Fishelson, and others.

Some Examples of ML Applications in Computational Biology

Dr. Yaron Kinar, Compugen

The field of computational biology offers a fertile ground for applying Machine Learning algorithms, looking for biological insights, as well as practical predictions, based on the vast amount of available experimental data. These applications may be used to assist the process of drug discovery. I will describe three such projects implemented in Compugen: a classifier for identifying alternatively spliced exons; another classifier for finding protein cleavage sites; and a Bayesian model for learning transcriptional networks. The talk will include a short biological introduction and descriptions of the applications.

Sparsity and Generalization in Machine Learning

Prof. Ron Meir, Technion

Supervised learning algorithms in Machine Learning aim at achieving low generalization error, based on a finite set of training examples. Generalization bounds developed in recent years possess sufficient flexibility to be used to construct effective algorithms, which guarantee good performance. We review recent bounds and show how to use them in order to obtain low dimensional kernel representations, focusing on both feature selection and sparse representation. Numerical studies show that algorithms devised in this manner can achieve state-of-the-art performance on many real world data sets. Special emphasis is placed on computationally effective algorithms based on convex programming.

Joint work with Dori Peleg.

On the Application of Machine Learning in Simulation-based Verification Systems

Dr. Shai Fine, IBM Haifa Research Lab

Functional verification is a process that ensures that a design conforms to its specification. It is widely acknowledged as the bottleneck of the hardware design cycle. At present, up to 70% of design development time and resources are spent on functional verification.

The increasing complexity of hardware designs, combined with shorter time-to-market requirements, raise the need to develop new techniques and methodologies to help verification teams achieve their goals quickly and with limited resources. To date, the leading techniques for functional verification are formal and simulation-based verification. While formal verification techniques are very powerful, their computational complexity make them applicable only for small units. Therefore, simulation-based verification is the main functional verification vehicle for large and complex designs.

In simulation-based verification, the design is fed with stimuli designed to trigger architecture and micro-architecture events in the design. The design is then simulated and its behavior is observed to verify that it conforms to the specification.

I will provide a short introduction to modern simulation-based verification techniques, and describe the application of machine learning to meet some of the major challenges in this realm.

Information Theoretic Algorithms for Dimensionality Reduction

Prof. Naftali Tishby, School of Computer Science and Engineering and Interdisciplinary Center for Neural Computation, The Hebrew University, Jerusalem

The Maximum Entropy framework has been around for a very long time, as a principled approach to distributional inference under expectation constraints. It turns out that for classification and other learning tasks, mutual information is the proper functional to optimize, and it provides interesting inference rules for supervised learning under various assumptions.

I will review several new algorithms that result from this approach with some applications to machine learning. - what's Machine Learning got to do with it?

Oren Glickman, is a leading online comparison shopping search engine and service. This presentation will describe a few concrete problems has encountered for which Machine Learning (ML) techniques were successfully applied. This includes an interesting range of methods such as supervised learning, clustering, and Reinforcement Learning.

Unsupervised Learning of Pair-wise Lexical Entailment

Dr. Ido Dagan, Bar-Ilan University

This talk discusses the problem of learning pair–wise entailment relationships between words-a kind of a lexical subsumption relationship.

In general terms, a word w entails a word v if the meaning of v can be inferred from the meaning of w, such that substituting v by w in some sentences would still imply the original meaning that was expressed using v. For example, the words "company" and "firm" entail each other (in certain contexts), while "bank" entails "company", but not vice versa. This type of lexical relationship is useful for information access applications, such as information retrieval, question answering, and information extraction, in order to identify different ways in which some target information may be expressed.

We present an unsupervised approach for learning lexical entailment relationships based on distributional feature vectors, similar to distributional object representations in clustering. Our work addresses two major issues. The first is designing a proper feature weighting function, such that the most prominent features of a word would indeed be highly indicative of its meaning, which also yields substantial feature reduction. The second issue is a proper scheme for comparing two feature vectors, based on a feature inclusion test that is augmented using web information to overcome feature data sparseness. Empirical tests show substantial improvements relative to classical word similarity approaches in the field.

Learning to Estimate Query Difficulty

Dr. Elad Yom-Tov, IBM Haifa Research Lab

We will present novel learning methods for estimating the quality of results returned by a search engine in response to a query. The estimation is based on agreement between the top results of the full query and the top results of its sub-queries.

We will demonstrate the usefulness of quality estimation for several applications, including improvement of retrieval, detecting queries for which no relevant content exists in the document collection, and distributed information retrieval. Experiments on TREC data demonstrate the robustness and the effectiveness of our learning algorithms.

Regret Minimization Algorithms

Prof. Yishay Mansour, Tel Aviv University

The regret minimization model considers an online agent that, at each time step, selects an action among N possible actions. There are a few natural models depending on the information the agent observes. In the "full information" model, the agent observes, at each time step, the loss of every action, While in the "partial information" (bandit) model, it observes only the loss of the action it selected.

In general, the agent would like to maximize its benefit. There are a few leading performance measures that "regret minimization" studies. External regret compares the performance of the agent to the performance of the best single actions in hindsight. Internal regret compares the loss of the agent to the loss of a modified online algorithm, which consistently replaces one action by another.

The main theme of the classical regret minimization results is that one can achieve vanishing average regret, without any assumptions concerning the generation of the reward sequence. The talk outlines a few recent simple online algorithms for the various regret minimization settings (both the full and partial information models and both external and internal regret). The results include both reductions between the models and the regret types as well as a new and novel external regret algorithm.

The talk is based on a number of joint works with Baruch Awerbuch, Avrim Blum, Nicolo Cesa-Bianchi, and Gilles Stoltz.


Related Seminar Links
Visitors information  
Machine Learning in HRL  

    About IBMPrivacyContact