Uncertainty in Deep Learning
Section: Full talk Technical level: Intermediate
How do you deal with uncertainty when making decisions? Presumably, you would collect more information to reduce the uncertainty before making a decision. Now, think about the outputs of deep learning models which can be used to make automated decisions. How will you get uncertainty estimates for these outputs? In this talk, we will focus on quantifying model uncertainty based on recent research and discuss potential applications of the uncertainty estimates.
Problem: Deep Neural Networks are powerful function approximators. However, it is well known that they are prone to overfitting and can be overconfident about their decisions.
- examples of overfitting
Solution: Bayesian approach: Discuss recent research that provides model uncertainty estimates of deep neural networks.
- bayesian neural nets
- weight uncertainty
- variational approximation
Applications of uncertainty estimates: Discuss the practical utility of the uncertainty estimates
- active learning
- exploration vs. exploitation
Madhu Gopinathan is currently Vice President, Data Science at MakeMyTrip, India’s leading online travel company. He has extensive experience in developing large scale systems using machine learning and natural language processing in both the Bay Area, USA and India. He has been granted multiple US patents, holds a PhD in the mathematical modeling of systems from the Indian Institute of Science, Bangalore and an MS in computer science from the University of Florida, Gainesville, USA.
Introduction to Game Training using Deep RL
From AlgphaGo to VizDoom, Deep Reinforcement Learning has revolutionarized the way in which we learn game environments. Especially for people playing CS and Dota in Colleges, having a smarter bot by thier side can help them sleep well, while the bots are fighting against each other. In the talk, we will discussing intuition behind design of Alphago. We will also discuss, what makes Reinforcement … more