IoP Colloquium: Max Welling (IvI, UvA)
Physics for Deep Learning and Deep learning for Physics
Deep learning is quickly developing into the workhorse for machine learning and artificial intelligence more broadly. In recent years it has revolutionised speech recognition (a field that was stagnant for many years), image analysis, natural language processing, machine translation, information retrieval, and more recently medical image analysis, and even parts of science and physics. Its successes are fuelled by massive compute power (e.g. GPUs developed just for deep learning), massive amounts of data and massive investments from industry.
In this talk I will start with a gentle introduction to machine learning and deep learning. I will then discuss how ideas imported from physics and mathematics have started to influence this field more recently. For instance, I will show how symmetries described by Lie Groups have led to a deep understanding (pun intended) of the abstract representations formed by these networks, as well as improved their performance. As a second example I will discuss how we can describe Bayesian deep learning using a "free energy" and thus reveal interesting connections with statistical physics. I will end by illustrating how this approach can compress deep neural nets by a factor of 500 with almost no loss in accuracy.
About the Speaker
Max Welling (UvA) received his PhD in ’98 under supervision of Nobel laureate Prof. G. 't Hooft. He received grants from Google, Facebook, Yahoo, NSF, NIH, NWO and ONR-MURI among which an NSF career grant in 2005. He is in the board of the Data Science Research Center in Amsterdam, directs the Amsterdam Machine Learning Lab (AMLAB), and co-directs the Qualcomm-UvA deep learning lab (QUVA), the Bosch-UvA Deep Learning lab (DELTA) and the AML4Health Lab.
If you missed this colloquium or would like to revisit it, the slides are available below.