Animashree Anandkumar – Bridging the gap between artificial and human intelligence: Role of Feedback
Abstract: Deep learning has yielded impressive performance over the last few years. However, it is no match to human perception and reasoning. Recurrent feedback in the human brain is shown to be critical for robust perception. Feedback is able to correct the potential errors made by the feed-forward pathways using an internal generative model of the world. The Bayesian brain hypothesis states that the human brain is carrying out Bayesian inference to obtain a consistent prediction. Deriving inspiration from this, we augment any existing neural network with feedback (NN-F). The feedback connections form a deconvolutional generative model that is Bayes-consistent with the given feed-forward neural network. We demonstrate inherent robustness in NN-F without any access to noisy examples, and further enhanced robustness when noisy examples are available.
In the second part of my talk, I will discuss some tools to analyze feedback mathematically, drawing from linear control theory. I will discuss a new reinforcement learning method that is able to achieve surprisingly low regret (logarithmic) using a combination of long-term and online learning. I will also discuss robust learning methods that can maintain safety and stability criteria, essential for feedback control systems. I will then discuss ways to bridge mathematical theory with real-world requirements.
Bio: Anima Anandkumar is a Bren Professor at Caltech and Director of ML Research at NVIDIA. She was previously a Principal Scientist at Amazon Web Services. She has received several honors such as Alfred. P. Sloan Fellowship, NSF Career Award, Young investigator awards from DoD, and Faculty Fellowships from Microsoft, Google, Facebook, and Adobe. She is part of the World Economic Forum’s Expert Network. She is passionate about designing principled AI algorithms and applying them in interdisciplinary applications. Her research focus is on unsupervised AI, optimization, and tensor methods.