UNIVERSITY OF HERTFORDSHIRE
             COMPUTER SCIENCE RESEARCH COLLOQUIUM

                            presents

             "Deeply interactive learning systems"

                      Dr. David Spivak
        (The Topos Institute, Berkeley and Department of Mathematics, MIT)

                     17 November 2021 (Wednesday)
                      17:00 -18:00 (note anomalous time)

                  Everyone is Welcome to Attend

                          (over Zoom)

Abstract:

We discuss a category-theoretic analysis of the mathematical structure of deep learning: the network of parameterized functions through which values are propagated forward and change is propagated backward. The abstract mathematical characterization of this structure shows a strong analogy with interacting open dynamical systems: machines wired together to form more complex machines, e.g. transistors wired together to form logic gates, adder circuits, and eventually computers and the internet. The left ⇄ right passage of information in a neural network corresponds to the inside ⇄ outside processing of information in a system of systems.

We propose a common generalization of these two structures, which we call "deeply interactive learning systems". To DNNs, we add the ability to include circuitry such as logic gates or feedback control—various sorts of memory and peer-to-peer interaction—inside each layer of the network, without changing the basic training structure. To interacting dynamical systems, we add the ability for the wiring pattern to change in time, moving from something like fixed integrated circuits to something like FPGAs.

This story came from a category-theoretic analysis of deep learning and dynamical systems, but I won't assume you know category theory for this talk.

---------------------------------------------------
Hertfordshire Computer Science Research Colloquium
http://cs-colloq.cs.herts.ac.uk