Biologically plausible learning in visco-elastic materials using predictive coding. Energy-based model approach to learning based on contrastive hebbian learning (CHL) in predictive coding networks.
Recent evidence has shown that computation can be achieved in visco-elastic materials using a learning rule inspired by a biologically-plausible energy-based algorithm known as equilibrium propagation (Stern et al., 2021). However, we argue that this learning rule is limited in biological plausibility given the need for storing memory of previous network states, as well as the requirement for external supervision in defining the targets during training.
We therefore use recent work by Millidge et al. (2022) in energy-based models to derive a new learning rule based on contrastive hebbian learning in predictive coding networks that is more biologically-plausible, and could conceivably be conducted fully-autonomously in a real physical system without the need for external supervision. We show that this learning rule, which we term generalised prospective configuration (GPC), is capable of learning desired network responses given target inputs. We argue that this learning rule provides tentative evidence for a simplified model for the way in which computation could plausibly emerge in organic materials.
GPC = Generalised Prospective Configuration (based on CHL in Predictive Coding Networks)
We attempt to replicate the initial task undertaken by Stern et al. (2021) using their Generalised Coupled Learning (GCL) rule based on Equilibrium Propagation: that of learning allosteric responses within a network.
The concept of allostery arises from biology, referring to the regulatory phenomenon by which the binding of a molecule locally to a site on a macromolecule (commonly a protein) can have an effect on the activity of that macromolecule at a distal site. This is usually caused by the coupling of conformational changes in the protein (Motlagh et al., 2014).
- GCL_allosteric_responses.ipynb replicates the work of Stern et al. (2021)
- GPC_allosteric_responses.ipynb implements our own learning rule (GPC)
Cite as:
P. Z. Collis, “Computation in Physical Media: Investigating Biologically-Plausible Learning in Visco-Elastic Materials via Predictive Coding,” M.S. thesis, Dept. of Eng. and Inf., Univ. of Sussex, Falmer, Brighton, 2022. [online].