UW - MSR Machine Learning workshop 2015 - Session 1

Опубликовано 27 июня 2016, 21:22

9:30 Opening Remarks - Ran Gilad-Bachrach 9:45 Toward Scalable Visual Recognition - Ali Farhadi Recognition is graduating from labs to real-world applications. While it is encouraging to see its potential being tapped, it brings forth a fundamental challenge: scalability. How can we learn a model for any concept that exhaustively covers all its appearance variations, while requiring minimal or no human supervision for compiling the vocabulary of visual variance, gathering the training images and annotations, and learning the models? In this work, I will introduce a fully-automated approach for learning extensive models for a wide range of variations (e.g. actions, interactions, attributes and beyond) within any concept. I will also showcase novel applications that our fully automated system enables including scalable image segmentation and visual verification. 10:10 Advances in Anomaly Detection - Tom Dietterich Our team at Oregon State University has developed several new algorithms for anomaly detection. These are based on two main principles: "anomaly detection by underfitting" and "anomaly detection by overfitting". In the underfitting approach, a model is fit to the data and points that do not fit well (e.g., that have low estimated probability density) are flagged as anomalies. In the overfitting approach, we transform the data to create learning problems in which there should be no signal or pattern and then apply machine learning algorithms to fit this data. If the algorithm finds a pattern, this is due to overfitting, and points belonging to the (false) pattern are likely to be anomalies. This talk will present these algorithms and also describe our benchmarking framework, which allows us to measure and compare the performance of different anomaly detection algorithms. I will also describe the results of a red-team experiment conducted under the DARPA ADAMS program in which our anomaly detection methods are showing excellent performance. 10:35 Spotlight: Blitz: A Principled Meta-Algorithm for Scaling Sparse Optimization - Tyler Johnson By reducing optimization to a sequence of small subproblems, working set methods achieve some of the fastest convergence times for many challenging problems. Despite their excellent performance, theoretical understanding of working set methods is limited, and implementations often resort to heuristics to determine subproblem size, makeup, and stopping criteria. We propose Blitz, a fast working set algorithm accompanied by useful guarantees. Making no assumptions on data, our theory relates subproblem size to progress toward convergence. This result motivates methods for optimizing algorithmic parameters and eliminating irrelevant variables as iterations progress. Applied to L1-regularized learning, these insights translate to state-of-the-art performance, as Blitz convincingly outperforms existing solvers in sequential, limited-memory, and distributed settings. Blitz is not specific to L1-regularized learning, making the algorithm relevant to many applications involving sparsity or constraints. 10:40Spotlight: Reactive learning - Christopher Lin One of the most popular uses of crowdsourcing is to provide training data for supervised machine learning algorithms. When we wish to train the best classifier possible at the lowest cost, active learning algorithms can intelligently pick new examples for the crowd to label. However, because the labels are noisy, algorithms should also be able to pick examples to relabel, in order to decrease the overall noise of the training data. In our study of reactive learning, we seek to understand the difference in marginal value between decreasing the noise of a training set, via relabeling, and increasing the diversity of the training set, via labeling new examples. The ultimate goal is an end-to-end decision theoretic system that, given a new learning problem, can dynamically make these optimal tradeoffs in pursuit of the most accurate classifier obtainable for a given labeling budget.