Jon will go over his website, nets-vs-automata.net, which is a distributed study pitting the statistical pattern recognition power of neural networks against the deterministic complexity of cellular automata. All experiments for this study are run by users in their browser, using tensorflow.js, and results are aggregated on the backend and are available for analysis. The main goal of the study is to map the six-dimensional performance metric landscape (e.g. accuracy, precision, recall, etc.) over the thirteen-dimensional training parameter input space (e.g. number of neural net hidden layers, training examples, epochs, etc.). There are over 393 trillion experiments that can be run, and each helps map out a point on this "surface". In many situations the nets don't perform well, while in other cases they become highly accurate at predicting future states of the cellular automata. Jon will give you a little intuition into when the nets perform well and when they don't using the nets-vs-automata analysis page. Attendees can also read the following academic paper related to the presentation ahead of time, if they are very interested in the topic: https://arxiv.org/abs/1809.02942
Jon Khanlian is an actuary, programmer, and filmmaker interested in machine learning. He made a movie that is somewhat related to this talk called "Digital Physics", which is available to watch on Tubi, Apple, Amazon, and other platforms. Other recent projects include scriptmonger.org, a free screenplay writing software, and PathToTheCup.com, a NY Red Bulls soccer blog. Disclaimer: He has only dabbled in Haskell and Lisp, but finds the benefits of functional programming compelling.
