Tuesday, October 1, 2013 - 12:30
Local Low-Rank Matrix Approximation
Tuesday, July 16, 2013 - 12:30
Method-of-Moment Algorithms for Learning Bayesian Networks
Tuesday, June 4, 2013 - 12:30
Which Supervised Learning Method Works Best for What? An Empirical Comparison of Learning Methods and Metrics
Decision trees can be intelligible, but do they perform well enough that you should use them? Have SVMs replaced neural nets, or are neural nets still best for regression and SVMs best for classification? Boosting maximizes a margin similar to SVMs, but can boosting compete with SVMs? If it does, is it better to boost weak models or to boost stronger models? Bagging is easier than boosting - how well does it stack up against boosting? Breiman said Random Forests are better than bagging and as good as boosting. Was he right? And what about old friends like logistic regression, KNN, and naive Bayes? Should they be put out to pasture, or do they fill important niches?
In this talk we'll compare the performance of ten supervised learning methods on nine criteria: Accuracy, F-score, Lift, Precision/Recall Break-Even Point, Area under the ROC, Average Precision, Squared Error, Cross-Entropy, and Probability Calibration. The results show that no one learning method does it all, but some methods can be 'repaired' to do very well across all performance metrics. In particular, we show how to obtain good probabilities from max-margin methods such as SVMs and boosting via Platt's Method and Isotonic Regression. We also describe a meta-ensemble method that combines select models from these ten learning methods to yield even better performance than any of the individual learning methods. Although these ensembles perform extremely well, they are too complex for some real-world applications. We'll describe a model compression method that tries to fix that. Finally, if time permits, we'll discuss how the nine performance metrics relate to each other, and on which of the metrics you probably should (or shouldn't) depend.
Tuesday, May 28, 2013 - 12:30
Reproducibility and Probabilistic Tsunami Hazard Assessment
I will discuss two somewhat independent topics. The first is the importance of reproducibility in computational research and the need for higher standards and better tools to help insure that published results can be reproduced (preferably by other researchers, but at least by the authors!). I will mention some of the tools already available to help with this.
Tsunami modeling is a motivating example since the results of numerical simulations are often used to inform public policy decisions that can have significant financial and safety implications. I will also briefly describe some of the techniques used to perform probabilistic hazard assessment and to solve the inverse problems necessary to do real-time forecasting, in hopes of identifying overlapping interests with the machine learning community.
Tuesday, May 21, 2013 - 12:30
Local Privacy, Minimax Rates, and Learning
We study statistical estimation minimization problems under a privacy model in which the data is kept confidential even from the learner. In this local privacy framework, we establish sharp upper and lower bounds on the convergence rates of statistical inference procedures. As a consequence, we exhibit a precise tradeoff between the amount of privacy the data preserves and the utility, as measured by convergence rate, of any statistical estimator or learning procedure.
Joint work with Michael Jordan and Martin Wainwright.
Tuesday, May 14, 2013 - 12:30
Recursive Deep Learning for Modeling Semantic Compositionality
Compositional and recursive structure is commonly found in different modalities, including natural language sentences and scene images. I will introduce several recursive deep learning models that, unlike standard deep learning methods can learn compositional meaning vector representations for phrases, sentences and images. These recursive neural network based models obtain state-of-the-art performance on a variety of syntactic and semantic language tasks such as parsing, paraphrase detection, relation classification and sentiment analysis.
Besides the good performance, the models capture interesting phenomena in language such as compositionality. For instance the models learn different types of high level negation and how it can change the meaning of longer phrases with many positive words. They can learn that the sentiment following a "but" usually dominates that of phrases preceding the "but." Furthermore, unlike many other machine learning approaches that rely on human designed feature sets, features are learned as part of the model.
I will focus on two recent results: Pushing the performance of the Stanford parser by 3.8% and improving its speed by 20% using syntactically untied RNNs and a recursive neural tensor network trained on a novel sentiment treebank.
Tuesday, May 7, 2013 - 12:30
Validating Network Classifiers and Pricing Information
Validating Network Classifiers: Networks are fundamental to our lives, from the network of gene interactions that shapes our bodies to the social networks that can eat up the hours of our lives. :) Collective classification uses network structure to predict node information. For example, if your friends all like jazz, are you likely to as well? Since networks grow by adding nodes based on the nodes already in the network, nodes are not drawn i.i.d. This makes trouble for most machine learning approaches to validation of classifier performance. We will discuss a method to validate network classifiers that is based on understanding how the network grows.
Pricing Information: In auctions for online advertising, data providers tell advertisers which users are the best bets for their ads. So advertisers buy a combination of information (from data providers) and advertising space (from publishers like Yahoo). How much should advertisers pay for each? Let's have a hands-on experience to find out.
Tuesday, April 30, 2013 - 12:30
Interpretable patient-level predictive models
This talk presents statistical methods which generate patent-level predictions that are both accurate and highly interpretable to healthcare providers and patients. In this context, an interpretable model should be able to pinpoint exactly why a particular prediction was made, and provide the reason in a clear and natural way. The talk begins by introducing the Hierarchical Association Rule Model (HARM) which sequentially predicts a patient's possible future medical conditions given the patient's current and past history of reported conditions. The core of our technique is a Bayesian hierarchical model for selecting predictive association rules (such as ``dyspepsia and epigastric pain imply heartburn") from a large set of candidate rules. We next present a model for traditional classification problems based on decision lists, which consist of a series of if...then... statements (for example, if high blood pressure, then stroke). Decision lists discretize the high-dimensional, multivariate feature space into a series of simple, readily interpretable decision statements. Our Bayesian framework, known as the Bayesian List Machine (BLM), introduces a formal relationship between sparsity and interpretability through a prior structure over lists. We compare our model with the CHADS2 score, actively used in clinical practice for estimating the risk of stroke in patients that have atrial fibrillation. Our model is as interpretable as CHADS2, but more accurate.
This is collaborative work with Cynthia Rudin, Ben Letham, and David Madigan.