IBM R&D Labs in Israel

Abstract: When looking for similar objected, like images and documents, and especially when querying a large remote data-base for similar objects, it is often useful to construct short similarity-preserving binary hashes.  That is, to map each image or document to a short bit strings such that similar objects have similar bit strings.  Such a mapping lies at the root of nearest neighbor search methods such as Locality Sensitive Hashing (LSH) and is recently gaining popularity in a variety of vision, image retrieval and document retrieval applications.  In this talk I will demonstrate, both theoretically and empirically, that even for symmetric and well behaved similarity measures, much could be gained by using two different hash functions---one for hashing objects in the database and an entirely different hash function for the queries.  Such asymmetric hashings can allow to significantly shorter bit strings and more accurate retrieval.

Abstract: Information systems are becoming increasingly reliant on statistical inference and learning to render all sorts of decisions, including the issuing of bank loans, the targeting of advertising, and the provision of health care. This growing use of automated decision-making has sparked heated debate among philosophers, policy-makers, and lawyers, with critics voicing concerns with bias and discrimination. Bias against some specific groups may be ameliorated by attempting to make the automated decision-maker blind to some attributes, but this is difficult, as many attributes may be correlated with the particular one. The basic aim then is to make fair decisions, i.e., ones that are not unduly biased for or against specific subgroups in the population. We formulate fairness as an optimization problem of finding a good representation of the data with two competing goals: to encode the data as well as possible, while simultaneously obfuscating any information about membership in the specific group. This is a computationally challenging objective, with links to several problems and approaches, including anonymity and the information bottleneck. I will present an initial model towards this goal, and show that it allows trade-offs between the system desiderata. I will also describe a direction we are currently exploring to extend the approach, to allow richer nonlinear representations.
Joint work with Cynthia Dwork, Moritz Hardt, Toni Pitassi, Omer Reingold, Kevin Swersky, Yu Wu, and Max Welling.

Abstract: We present the Discriminative Ferns Ensemble (DFE) classifier for efficient visual object recognition. The classifier architecture is designed to optimize both classification speed and accuracy when a large training set is available. The proposed framework is applied to the problem of hand pose recognition in depth and infra-red images, using a very large training set. Both the accuracy and the classification time obtained are considerably superior to relevant competing methods, allowing one to reach accuracy targets with run times orders of magnitude faster than the competition. We also show empirically that using DFE, we can significantly reduce classification time by increasing training sample size for a fixed target accuracy. The result classifier is now used for hand pose classification in Microsoft Xbox One.
Joint work with Ben Klein, Alon Vinnikov, Aharon Bar-Hillel, Daniel Freedman, Simon Stachniak.

Abstract: The goal of discriminative learning is to train a system to optimize the evaluation metric used to measure its performance. In binary classification one typically tries to minimizes the 0-1 loss, but in more complex prediction problems, each task has its own evaluation metric, such as NDCG in search engines, word error rate in speech recognition, or the BLEU score in machine translation. In the talk, I will discuss how current models of structured prediction such as structured support machine (SVMs) and conditional random fields (CRF) handle the evaluation metric optimization, and will present two algorithms that are designed to the optimize structured evaluation metrics.

Abstract: In this talk we will presents several applications of machine learning to various fields of design in an attempt to challenge the machine-learning community to expand towards non-traditional domains. Through investigations in architecture, social media, graphical design, industrial design and political art, we will show that machine learning can evolve to become a powerful tool in augmenting artistic statements and enhancing product usability and personalization.

Abstract: Learning the association between observed variables and future trajectories of continuous-time stochastic processes is a fundamental task in dynamic modeling. Often the dynamics are non-homogeneous and involve a large number of interacting components. In this work we introduce a conditional probabilistic model that captures such dynamics while maintaining scalability and providing an explicit way to express the interrelation between the system components. The principal idea is a factorization of the model into two distinct elements, one that depends only on time, and the other depends on the system configuration. We develop a learning procedure given either full or point observations and test it on simulated data. We apply the proposed modeling scheme to study EuResist, a cohort of HIV patients who underwent therapy, and demonstrate that the factorization helps to shed light on the dynamics of HIV.
Joint work with Omer Weissbrod and Elad Eban.

Abstract: Recent years have seen a dramatic resurgence of interest in deep learning systems (e.g. neural networks), capable of compactly representing highly non-linear and complex predictors. While providing groundbreaking empirical results on several practical problems, these methods often require considerable engineering effort, have many parameters to tune, and are very heuristic in nature (in particular, they come with no formal guarantees). In this talk, I'll survey these issues, and describe a recent attempt to address them in the context of polynomial predictors.