JSALT 2015 -- Week 5 Informal Seminar with Les Atlas
Tuesday, Aug. 4:
- 11am-12pm Informal seminar [EEB 303]
Active Learning, Redux
Active learning is a subfield of modern machine learning. If a learning algorithm is allowed to choose which data from which it learns and/or improves its performance, it can be shown, analytically, and in many applications, to achieve better accuracy. Part of the original inspiration for active learning came from how human babies, likely the best learning machines on the planet, acquire speech and language. More is now known about that early learning process. Also, active learning is similar, though not identical, to data subset selection, e.g. [Wei, Liu, Kirchhoff, Bilmes, Proc. IEEE ICASSP, 2014]. That potential connection suggests that performance guarantees, previously missing, might be possible for a modernized version of active learning. Lastly, our original 1987-94 vision for active learning was for a system which could use what is now called crowdsourcing to provide systems which improved as they interacted with a large community of users. Language learning concepts like Duolingo might finally offer the potential platform for this original vision, along with Amazon Echo, Microsoft Cortana, and Google Voice.
Les Atlas, received his Ph.D. degrees in Electrical Engineering from Stanford University in 1984. He then joined the University of Washington where he is a Professor of Electrical Engineering and Adjunct Professor of Computer Science and Engineering. His research is in digital signal processing, with specializations in audio and acoustic analysis, harmonic analysis and modulation representations, statistical signal detection, recognition, and coding. His early academic work has had recent practical impact: For example, he co-authored the first publication on convolutional neural networks for temporal signals [Homma, Atlas and Marks, “An artificial neural network for spatio-temporal bipolar patterns: Application to phoneme classification,” Proc. NIPS, 1988.] demonstrating accurate performance on speech phoneme classification. Also, one of his joint publications [Cohn, Atlas, and Ladner, “Improving generalization with active learning,” Machine Learning, 1994] helped initiated the machine learning fields of active learning and selective sampling. He is Co-Chair of the current 2015 Jelinek Workshop on Machine Learning for Speech and Language, which has brought many top visitors to the University of Washington to collaborate with the excellent students and faculty here.