arXiv subjects
Publications and source records attributed to Yoav Freund.
We explore a novel approach to semi-supervised learning. This approach is contrary to the common approach in that the unlabeled examples serve to "muffle," rather than enhance, the guidance provided by the labeled examples. We provide several variants of the basic algorithm and show experimentally that they can achieve significantly higher AUC than boosted trees, random forests and logistic regression when unlabeled examples are available.
We present and empirically evaluate an efficient algorithm that learns to aggregate the predictions of an ensemble of binary classifiers. The algorithm uses the structure of the ensemble predictions on unlabeled data to yield significant performance improvements. It does this without making assumptions on the structure or origin of the ensemble, without parameters, and as scalably as linear learning. We empirically demonstrate these performance gains with random forests.
We develop a worst-case analysis of aggregation of classifier ensembles for binary classification. The task of predicting to minimize error is formulated as a game played over a given set of unlabeled data (a transductive setting), where prior label information is encoded as constraints on the game. The minimax solution of this game identifies cases where a weighted combination of the classifiers can perform significantly better than any single classifier.
We consider a situation in which we see samples in $\mathbb{R}^d$ drawn i.i.d. from some distribution with mean zero and unknown covariance A. We wish to compute the top eigenvector of A in an incremental fashion - with an algorithm that maintains an estimate of the top eigenvector in O(d) space, and incrementally adjusts the estimate with each new data point that arrives. Two classical such schemes are due to Krasulina (1969) and Oja (1983). We give finite-sample convergence rates for both.
We consider using an ensemble of binary classifiers for transductive prediction, when unlabeled test data are known in advance. We derive minimax optimal rules for confidence-rated prediction in this setting. By using PAC-Bayes analysis on these rules, we obtain data-dependent performance guarantees without distributional assumptions on the data. Our analysis techniques are readily extended to a setting in which the predictor is allowed to abstain.
Handwriting is a natural and versatile method for human-computer interaction, especially on small mobile devices such as smart phones. However, as handwriting varies significantly from person to person, it is difficult to design handwriting recognizers that perform well for all users. A natural solution is to use machine learning to adapt the recognizer to the user. One complicating factor is that, as the computer adapts to the user, the user also adapts to the computer and probably changes their handwriting. This paper investigates the dynamics of co-adaptation, a process in which both the computer and the user are adapting their behaviors in order to improve the speed and accuracy of the communication through handwriting. We devised an information-theoretic framework for quantifying the efficiency of a handwriting system where the system includes both the user and the computer. Using this framework, we analyzed data collected from an adaptive handwriting recognition system and characterized the impact of machine adaptation and of human adaptation. We found that both machine adaptation and human adaptation have significant impact on the input rate and must be considered together in order to improve the efficiency of the system as a whole.
The sensitivity of Adaboost to random label noise is a well-studied problem. LogitBoost, BrownBoost and RobustBoost are boosting algorithms claimed to be less sensitive to noise than AdaBoost. We present the results of experiments evaluating these algorithms on both synthetic and real datasets. We compare the performance on each of datasets when the labels are corrupted by different levels of independent label noise. In presence of random label noise, we found that BrownBoost and RobustBoost perform significantly better than AdaBoost and LogitBoost, while the difference between each pair of algorithms is insignificant. We provide an explanation for the difference based on the margin distributions of the algorithms.
We study the tracking problem, namely, estimating the hidden state of an object over time, from unreliable and noisy measurements. The standard framework for the tracking problem is the generative framework, which is the basis of solutions such as the Bayesian algorithm and its approximation, the particle filters. However, these solutions can be very sensitive to model mismatches. In this paper, motivated by online learning, we introduce a new framework for tracking. We provide an efficient tracking algorithm for this framework. We provide experimental results comparing our algorithm to the Bayesian algorithm on simulated data. Our experiments show that when there are slight model mismatches, our algorithm outperforms the Bayesian algorithm.
High-throughput screening (HTS) using model organisms is a promising method to identify a small number of genes or drugs potentially relevant to human biology or disease. In HTS experiments, robots and computers do a significant portion of the experimental work. However, one remaining major bottleneck is the manual analysis of experimental results, which is commonly in the form of microscopy images. This manual inspection is labor intensive, slow and subjective. Here we report our progress towards applying computer vision and machine learning methods to analyze HTS experiments that use Caenorhabditis elegans (C. elegans) worms grown on agar. Our main contribution is a robust segmentation algorithm for separating the worms from the background using brightfield images. We also show that by combining the output of this segmentation algorithm with an algorithm to detect the fluorescent dye, Nile Red, we can reliably distinguish different fluorescence-based phenotypes even though the visual differences are subtle. The accuracy of our method is similar to that of expert human analysts. This new capability is a significant step towards fully automated HTS experiments using C. elegans.
We present a novel particle filtering algorithm for tracking a moving sound source using a microphone array. If there are N microphones in the array, we track all $N \choose 2$ delays with a single particle filter over time. Since it is known that tracking in high dimensions is rife with difficulties, we instead integrate into our particle filter a model of the low dimensional manifold that these delays lie on. Our manifold model is based off of work on modeling low dimensional manifolds via random projection trees [1]. In addition, we also introduce a new weighting scheme to our particle filtering algorithm based on recent advancements in online learning. We show that our novel TDOA tracking algorithm that integrates a manifold model can greatly outperform standard particle filters on this audio tracking task.
We study the tracking problem, namely, estimating the hidden state of an object over time, from unreliable and noisy measurements. The standard framework for the tracking problem is the generative framework, which is the basis of solutions such as the Bayesian algorithm and its approximation, the particle filters. However, the problem with these solutions is that they are very sensitive to model mismatches. In this paper, motivated by online learning, we introduce a new framework -- an {\em explanatory} framework -- for tracking. We provide an efficient tracking algorithm for this framework. We provide experimental results comparing our algorithm to the Bayesian algorithm on simulated data. Our experiments show that when there are slight model mismatches, our algorithm vastly outperforms the Bayesian algorithm.
We study the problem of decision-theoretic online learning (DTOL). Motivated by practical applications, we focus on DTOL when the number of actions is very large. Previous algorithms for learning in this framework have a tunable learning rate parameter, and a barrier to using online-learning in practical applications is that it is not understood how to set this parameter optimally, particularly when the number of actions is large. In this paper, we offer a clean solution by proposing a novel and completely parameter-free algorithm for DTOL. We introduce a new notion of regret, which is more natural for applications with a large number of actions. We show that our algorithm achieves good performance with respect to this new notion of regret; in addition, it also achieves performance close to that of the best bounds achieved by previous algorithms with optimally-tuned parameters, according to previous notions of regret.
We present a new boosting algorithm, motivated by the large margins theory for boosting. We give experimental evidence that the new algorithm is significantly more robust against label noise than existing boosting algorithm.
We present a new online learning algorithm for cumulative discounted gain. This learning algorithm does not use exponential weights on the experts. Instead, it uses a weighting scheme that depends on the regret of the master algorithm relative to the experts. In particular, experts whose discounted cumulative gain is smaller (worse) than that of the master algorithm receive zero weight. We also sketch how a regret-based algorithm can be used as an alternative to Bayesian averaging in the context of inferring latent random variables.
A simple and computationally efficient scheme for tree-structured vector quantization is presented. Unlike previous methods, its quantization error depends only on the intrinsic dimension of the data distribution, rather than the apparent dimension of the space in which the data happen to lie.
We present MEDUSA, an integrative method for learning motif models of transcription factor binding sites by incorporating promoter sequence and gene expression data. We use a modern large-margin machine learning approach, based on boosting, to enable feature selection from the high-dimensional search space of candidate binding sequences while avoiding overfitting. At each iteration of the algorithm, MEDUSA builds a motif model whose presence in the promoter region of a gene, coupled with activity of a regulator in an experiment, is predictive of differential expression. In this way, we learn motifs that are functional and predictive of regulatory response rather than motifs that are simply overrepresented in promoter sequences. Moreover, MEDUSA produces a model of the transcriptional control logic that can predict the expression of any gene in the organism, given the sequence of the promoter region of the target gene and the expression state of a set of known or putative transcription factors and signaling molecules. Each motif model is either a $k$-length sequence, a dimer, or a PSSM that is built by agglomerative probabilistic clustering of sequences with similar boosting loss. By applying MEDUSA to a set of environmental stress response expression data in yeast, we learn motifs whose ability to predict differential expression of target genes outperforms motifs from the TRANSFAC dataset and from a previously published candidate set of PSSMs. We also show that MEDUSA retrieves many experimentally confirmed binding sites associated with environmental stress response from the literature.
We present a novel classification-based method for learning to predict gene regulatory response. Our approach is motivated by the hypothesis that in simple organisms such as Saccharomyces cerevisiae, we can learn a decision rule for predicting whether a gene is up- or down-regulated in a particular experiment based on (1) the presence of binding site subsequences (``motifs'') in the gene's regulatory region and (2) the expression levels of regulators such as transcription factors in the experiment (``parents''). Thus our learning task integrates two qualitatively different data sources: genome-wide cDNA microarray data across multiple perturbation and mutant experiments along with motif profile data from regulatory sequences. We convert the regression task of predicting real-valued gene expression measurement to a classification task of predicting +1 and -1 labels, corresponding to up- and down-regulation beyond the levels of biological and measurement noise in microarray measurements. The learning algorithm employed is boosting with a margin-based generalization of decision trees, alternating decision trees. This large-margin classifier is sufficiently flexible to allow complex logical functions, yet sufficiently simple to give insight into the combinatorial mechanisms of gene regulation. We observe encouraging prediction accuracy on experiments based on the Gasch S. cerevisiae dataset, and we show that we can accurately predict up- and down-regulation on held-out experiments. Our method thus provides predictive hypotheses, suggests biological experiments, and provides interpretable insight into the structure of genetic regulatory networks.