MBC IGERT Graduate Training Seminar Series - Andrew Mass and Samir Menon

Deep Neural Networks in Speech Recognition Presented by Andrew Maas, Stanford

Abstract:  Speech recognition is a well-formalized, data-rich task where computational approaches still cannot match human performance. I will describe two projects on large vocabulary continuous speech recognition for conversational speech. Using an HMM-based approach to speech recognition, my team and I performed an extensive set of experiments using deep neural networks (DNNs) as acoustic models to map small spans of acoustic data to phonetic sub-states. Our results suggest that some aspects of DNN architecture are important for overall system performance, but overall it is difficult to obtain system performance gains in an HMM-based system by DNN modifications alone. This finding leads to the second project, which abandons the traditional HMM-based framework for speech recognition and instead utilizes deep recurrent networks to directly map from acoustic inputs to text characters. We demonstrate the first set of results using such a simple approach to yield reasonable performance on the challenging Switchboard corpus. Our model uses a character-level language model in place of a word-level lexicon and language model. This further reduces system complexity and enables the model to transcribe out of vocabulary words.

Using Robotics to Study Human Motor Control Presented by Samir Menon, Stanford

Abstract:  Robotic control theory, when applied to human musculoskeletal models, provides us with techniques to mathematically formulate human neuromuscular control and predict its representation in the brain. In this talk, I will outline our work in applying robotics insights to develop a control formulation for musculoskeletal models of the human arm. These models help us design feasible and informative neuroscience experiments that study human manipulation. I will focus on one specific aspect of human manipulation, the ability to generalize manipulation skills across a variety of body postures. Our models predict that in order to achieve this generalization, the brain should segregate the notion of a manipulation task (control) from the process of distributing task errors across the body's many muscles and joints (coordination). I will discuss results from a human neuroimaging experiment that offers the ability to decorrelate control from coordination. Our results, consistent with theoretical predictions, suggest that correlates of motor control and coordination can be localized to different anatomical regions in the brain. These results set the stage for future experiments that test targeted control theory based hypotheses of human motor control.

Dinner will be provided at 6:30pm.

If you plan on attending dinner, please make sure to RSVP to lehope@stanford.edu by Tuesday, April 7