Integrated Calendar

We can store and retrieve specific patterns of activity in network models through synaptic plasticity mechanisms. When the synapses between cells in these models are bounded, then encoding a new pattern necessarily implies the partial erasure of previously stored ones. This overwriting or “palimpsest” property of networks with biologically constrained synapses has been studied intensively over the past 30 years. Most theoretical studies have focused on mechanisms for improving the memory capacity in such networks, which is starkly degraded through overwriting. However, there is another property of these memory systems which has not yet been fully explored. Namely, in the context of sensorydriven activity, ongoing learning can lead to the overwriting of some fraction of the synapses. This in turn leads to changes in the output of the network at any two distinct points in time, even if the input patterns have remained unchanged. This effect is reminiscent of the phenomenon of representational drift (RD), which has by now been wellestablished in the hippocampus, and other cortical areas. Recent experimental work has brought to light a number of puzzling findings regarding RD, which seem to defy simple explanation. These include the discovery that repetition rate can both reduce drift (in piriform cortex) and increase it (in hippocampus). In hippocampal place cells, RD has been shown to have differential effects on overall firing rates and spatial tuning. This suggests that there may be distinct underlying mechanisms. I will discuss how all of these findings are, in fact, consistent with the changes in activity observed in networks which store patterns through Hebbian plasticity. The fundamental assumption in such models is that memory storage is ongoing, and occurs between experimental sessions. The array of distinct and sometimes seemingly contradictory findings can be accounted for by differences in learning rates and correlations between input patterns.

Ordered mechanical systems typically have one or only a few stable rest configurations, and hence are not considered useful for encoding memory. Multistable and history-dependent responses usually emerge from quenched disorder, for example in amorphous solids or crumpled sheets. Inspired by the topological structure of frustrated artificial spin ices, we introduce an approach to design ordered, periodic mechanical metamaterials that exhibit an extensive set of spatially disordered states. We show how such systems exhibit non-Abelian and history-dependent responses, as their state can depend on the order in which external manipulations were applied. We demonstrate how this richness of the dynamics enables to recognize, from a static measurement of the final state, the sequence of operations that an extended system underwent. Thus, multistability and potential to perform computation emerge from geometric frustration in ordered mechanical lattices that create their own disorder.

In this seminar, two new models will be presented. The first new model is a first-principles description of the propagation of light in a cloud, based on a classical solution to Maxwell's equations rather than radiative transfer theory. The second new model is a fully three-dimensional, time-dependent model of the regions of possible sprite inception in the mesosphere, based on the classical method of images from electrostatics rather than finite differencing in space. The reason why each model is unique, the problems each model can solve, and the kinds of results each model can produce will be discussed.