High Energy Physics Seminar

After a halting start in the late 1980s and early 1990s, characterized by great skepticism from many high energy physicists, machine learning is now firmly established within our analysis toolkit. I argue that the most important thing to know about machine learning methods, in particular, those based on the use of highly non-linear functions such as neural networks, deep or otherwise, is what these methods approximate. In this talk, I begin with a description of some recent musings on  deep learning applications in high energy physics in which I emphasize the connection to Bayes theorem. I then consider what I, and a growing number of machine learning enthusiasts, consider the greatest deficiency of machine learning methods, namely, their neglect of uncertainty. I argue that Bayes theorem, again, has a role to play in this regard.