Phenomenology of Deep Learning

While machine learning methods are broadly used in physics, the flow of ideas in the opposite direction, i.e., the use of concepts and techniques from theoretical physics to understand modern deep learning, has only started to be explored. In the first part of this talk, I will review the basics of neural networks and discuss why commonly used complexity measures cannot explain the success of the so-called modern deep-learning regime. In the second part, I will motivate how a physics-based framework can provide a unique perspective to pressing questions in deep learning, and contribute to filling the gap between practical developments and theoretical foundations. To conclude, I will discuss some examples in which physics model-building tools can be used to study intriguing phenomena observed in deep neural networks.

Zoom: https://zoom.us/j/91390797603?pwd=bWdEdUJNYkhZQWZQRFdsTjlSQXk0Zz09