In-memory learning and forgetting in networks with self-adjusting edge conductances

Contrastive local learning networks (CLLNs) consist of a highly-connected network of variable-conductance edges, in one example, where each edge independently self-adjusts its conductance using a local learning rule in order for voltages on output nodes to be a desired function of voltages on the input nodes. In these dynamical systems, the conductances of the edges depend upon - and hold memory of - their initialization and the history of how training data are shown to the network. In effect, they perform in-memory analog learning for in-memory analog computational tasks such as nonlinear regression and classification -- the kind of functionality routinely gained by Artificial Neural Networks (ANNs), but now without a processor. In this talk, we consider the behavior of CLLNs alternately trained for two different functional tasks. The two functionalities are compatible and, in principle, can be simultaneously achieved. However, in practice, the network can forget a first task if trained for too long on a second task. We show that such loss of memory can arise from bias in the physical implementation of the learning rules. This can be analytically modeled, and it can be mitigated by a generally-applicable "overclamping" trick that we introduce.