Self-timed self-supervised learning

Life can be easier if one knows the structure of the world, for
instance, that a distant roar, a whiff of a heavy musky smell, and
black stripes on orange background are caused by a single physical
entity. Indeed, the question how such structures can be discovered by
the neural networks of the brain has challenged neuroscientists for
many decades. A key constraint is that central nervous systems must
learn about the structure of the world from observing correlations
within continuous streams of spikes that arrive from their sensory
peripheries. Recently, a novel family of unsupervised spiking neural
network models, self-supervised neural networks, have been shown
highly potent in discovering ensembles of recurring spike patterns
even when their individual occurrences within background noise were
rare and temporally asynchronous. In these models an
internal supervisory circuit drives learning within a layer of
processing neurons by providing teaching signals computed from the
processing layer's past activity. A central limitation of these models
is their reliance on the presence of an externally given trial
structure: The given end of a sensory episode prompts the supervisory
circuit to compute its teaching signals and initiate a learning step
within the processing layer. Here we develop a self-timed version of
self-supervised networks whose teaching circuit requires neither
external clock nor trial-end signals, but rather uses the processing
layer's activity to also decide the timing of learning steps. We
demonstrate that self-timed self-supervised networks match the
performance (convergence times) of the original trial-based learning
model, substantially broadening the range of settings to which this
approach is applicable. We explore the stability of the learning
dynamics arising from the interactions between synaptic plasticity and
homeostatic mechanisms. Our work provides a biologically plausible
unsupervised neural network model for multi-modal learning of
recurring spike patterns within parallel continuous sensory
streams.