Abstract:Theoretical models of neuronal function consider different mechanisms through which networks learn, classify and discern inputs. A central focus of these models is to understand how associations are established amongst neurons, in order to predict spiking patterns that are compatible with empirical observations. Although these models have led to major insights and advances, they still do not account for the astonishing velocity with which the brain solves certain problems and what lies behind its creativity, amongst others features. We examine two important components that may crucially aid comprehensive understanding of said neurodynamical processes. First, we argue that once presented with a problem, different putative solutions are generated in parallel by different groups or local neuronal complexes, with the subsequent stabilization and spread of the best solutions. Using mathematical models we show that this mechanism accelerates finding the right solutions. This formalism is analogous to standard replicator-mutator models of evolution where mutation is analogous to the probability of neuron state switching (on/off). The second factor that we incorporate is structural synaptic plasticity, i.e. the making of new and disbanding of old synapses, which we apply as a dynamical reorganization of synaptic connections. We show that Hebbian learning alone does not suffice to reach optimal solutions. However, combining it with parallel evaluation and structural plasticity opens up possibilities for efficient problem solving. In the resulting networks, topologies converge to subsets of fully connected components. Imposing costs on synapses reduces the connectivity, although the number of connected components remains robust. The average lifetime of synapses is longer for connections that are established early, and diminishes with synaptic cost.

The Physical Effects of Learning

Physical effects of learning

Physical networks become what they learn

Learning dynamical behaviors in physical systems

Microscopic imprints of learned solutions in adaptive resistor networks

Emergent learning in physical systems as feedback-based aging in a glassy landscape

Experimental demonstration of coupled learning in elastic networks

Learning and Structure of Neuronal Networks

Physical Neural Networks with Self-Learning Capabilities

On Physical Origins of Learning

Learning Interacting Theories from Data

Coupled Hebbian learning and evolutionary dynamics in a formal model for structural synaptic plasticity

Physical learning of power-efficient solutions

Physically optimizing inference

Learning about learning by many-body systems

Self-shaping dynamical systems and learning

Intelligent mechanical metamaterials towards learning static and dynamic behaviors

On instabilities in neural network-based physics simulators

On the weight dynamics of learning networks

A theoretical framework for learning through structural plasticity