Overview Lectures:

• Overview of speech recognition using deep architectures beyond HMMs
• Recent developments in deep learning: Architectures and Algorithms

Topics for Panel Discussion and Open Discussion:

• Why speech is hard from the machine learning perspective --- Complexity and variability along the full chain of speech at all levels
• Limitations of shallow learning in speech recognition and related sequential pattern recognition applications
• Generative models vs. discriminative models with deep structure
• Integrated generative/discriminative deep architecture
• Supervised vs. unsupervised learning and the relationship to generative/discriminative deep architectures
• Representing sequential patterns
• Robustness issues: What happens when the input is noisy
• Scalability issues: toy problems vs. real applications; what are the main obstacles for wide adoption
• Can the model effectively learn the multiple-level, internal representations from raw features: pixels for images vs. waveforms for speech?
• Deriving the linguistic hierarchy and data-driven speech units using deep learning
• Building common evaluation tasks to assess theories, models, architectures, algorithms, and feature designs
• Relationship to how brains process speech and other sequential patterns

Contact Us Terms of Use Trademarks Privacy Statement ©2010 Microsoft Corporation. All rights reserved.