Talks

Recently, we have shown that deep neural networks can be used to solve the Schrödinger Equation (https://journals.aps.org/pra/abstract/10.1103/PhysRevA.96.042113), classical spin models (https://doi.org/10.1103/PhysRevE.97.032119), and 2d-materials such as graphene and boron-nitride (https://doi.org/10.1016/j.commatsci.2018.03.005).

I will argue that these "deep models" can be used to simulate and understand matter in a way which was not previously possible. Specifically, I will show how our recently reported extensive deep neural networks (https://arxiv.org/abs/1708.06686) can be used to infer the properties of meso-scale materials based on training data generated from much smaller structural motifs (evaluated using electronic structure methods such as density functional theory). Extensive deep neural networks scale as O(N) and can be efficiently evaluated in parallel using petascale computational platforms.

Landauer's famous dictum that 'information is physical' has been enthusiastically taken on by a range of communities, with researchers in areas from quantum and unconventional computing to biology, psychology, and economics adopting the language of information processing. However, this rush to make all science about computing runs the risk of collapsing into triviality: if every physical process is computing, then to say that something performs computation gives no meaningful information about it, leaving computational language devoid of content. In this talk I will give an introduction to Abstraction/Representation Theory, a framework for representing both computing and physical science that allows us to draw a meaningful distinction between them. The use of AR theory - with its commuting-diagrammatic framework and associated algebra of representation - allows us to take significant steps towards giving a formal language and framework for the processes of science. I will show how AR theory represents this process (including the potential for automation), and the insights it gives into the usage and limits of computation as a formal process language for, and description of, physical sciences.

In this talk, I show how information theoretic concepts can be used to extend the scope of traditional Bayesianism. I will focus on the learning of indicative conditionals (“If A, then B”) and a Bayesian account of argumentation. We will see that there are also interesting connections to research done in the psychology of reasoning. The talk is partly based on the paper “Bayesian Argumentation and the Value of Logical Validity” (with Ben Eva, forthcoming in Psychological Review, http://philsci-archive.pitt.edu/14491/).

When applied to a physical system, the two main, established notions of information, Shannon Information and Algorithmic Information, explicitly neglect the mechanistic structure of the system under evaluation. Shannon information treats the system as a channel and quantifies correlations between the system’s inputs and outputs, or between its past and future states. Algorithmic information quantifies the length of the shortest program capable of reproducing the system’s outputs or dynamics. The goal in both cases is to predict the system’s behavior from the perspective of an extrinsic investigator. From the intrinsic perspective of the system itself, however, information must be physically instantiated to be causally relevant. For every ‘bit’, there must be some mechanism that is in one of two (or several) possible states, and which state it is in must matter to other mechanisms. In other words, the state must be “a difference that makes a difference” and implementation matters. By examining the informational and causal properties of artificial organisms (“animats”) controlled by small, adaptive neural networks (Markov Brains), I will discuss necessary requirements for intrinsic information, autonomy, and agency.