Synapseek: Meta-learning synaptic plasticity rules

Animals exhibit a remarkable ability to learn and remember new behaviors, skills, and associations throughout their lifetime. These capabilities are made possible thanks to a variety of changes in the brain throughout adulthood, regrouped under the term "plasticity". Some cells in the brain —neurons— and specifically changes in the connections between neurons, the synapses, were shown to be crucial for the formation, selection, and consolidation of memories from past experiences. These ongoing changes of synapses across time are called synaptic plasticity. Understanding how a myriad of biochemical processes operating at individual synapses can somehow work in concert to give rise to meaningful changes in behavior is a fascinating problem and an active area of research. However, the experimental search for the precise plasticity mechanisms at play in the brain is daunting, as it is difficult to control and observe synapses during learning. Theoretical approaches have thus been the default method to probe the plasticity-behavior connection. Such studies attempt to extract unifying principles across synapses and model all observed synaptic changes using plasticity rules: equations that govern the evolution of synaptic strengths across time in neuronal network models. These rules can use many relevant quantities to determine the magnitude of synaptic changes, such as the precise timings of pre- and postsynaptic action potentials, the recent neuronal activity levels, the state of neighboring synapses, etc. However, analytical studies rely heavily on human intuition and are forced to make simplifying assumptions about plasticity rules. In this thesis, we aim to assist and augment human intuition in this search for plasticity rules. We explore whether a numerical approach could automatically discover the plasticity rules that elicit desired behaviors in large networks of interconnected neurons. This approach is dubbed meta-learning synaptic plasticity: learning plasticity rules which themselves will make neuronal networks learn how to solve a desired task. We first write all the potential plasticity mechanisms to consider using a single expression with adjustable parameters. We then optimize these plasticity parameters using evolutionary strategies or Bayesian inference on tasks known to involve synaptic plasticity, such as familiarity detection and network stabilization. We show that these automated approaches are powerful tools, able to complement established analytical methods. By comprehensively screening plasticity rules at all synapse types in realistic, spiking neuronal network models, we discover entire sets of degenerate plausible plasticity rules that reliably elicit memory-related behaviors. Our approaches allow for more robust experimental predictions, by abstracting out the idiosyncrasies of individual plasticity rules, and provide fresh insights on synaptic plasticity in spiking network models.