Inferring plasticity rules from single-neuron spike trains using deep learning methods

Synaptic plasticity is a core basis for learning and adaptation. Determining how synapses are altered by local signals – the learning rules – is the hinge about which brain activity pivots. A large number of in vitro characterizations have focused on restricted sets of core properties, but it remains to be established which if any of the known learning rules is most consistent with changes in activity patterns in behaving animals. To address this question, we hypothesize that the correlation between features of the activity of a single post-synaptic neuron and subsequent changes of the representations could be used to detect the underlying learning rule. Because this correlation is expected to be diluted in the notoriously large variability of brain activity, we test here learning rule inference based on passive observations of single neurons using deep artificial neural networks. Using simulated data, we found that both transformers, temporal convolutional networks, and SVM could classify learning rules far above the chance level, with transformers achieving the best overall accuracy. This performance can be achieved despite the presence of noise and representational drift. We further investigated the features used by the algorithms to perform the classification and found the deep net used inner temporal differences of distinct learning rules to separate learning trajectories. We also find, however, that the classification accuracy is sensitive to alterations in network properties. Our work illustrates that distinct learning rules’ generate distinguishable trajectories of responses, but warns against using simulation-trained classifiers to infer learning rules from real data.