Memory by a thousand rules: Automated discovery of multi-type plasticity rules reveals variety and degeneracy at the heart of learning.

Synaptic plasticity is widely understood to be the basis of learning and memory, but the link between synaptic changes and network properties remains elusive. Here, we use automated search algorithms to obtain thousands of excitatory(E)-to-E, E-to-inhibitory(I), IE and II plasticity rules working in cooperation in recurrent spiking networks. We selected these rule quadruplets solely based on their ability to robustly stabilize the networks they were embedded in. Simple forms of memory such as novelty detection appeared to be nigh-unavoidable byproducts of this homeostasis criterion. We further inferred rules that reproduce more complex network functions seen experimentally such as contextual novelty and replay. Co-activity was crucial: most rules were unstable in isolation. Overall, we demonstrate the viability of automated discovery algorithms for synaptic plasticity at scale, and show that relatively simple, unsupervised rules can elicit complex network functions through cooperation.