Demonstration of Synaptic Behaviour in a Spintronic Device for On-Chip Learning in Long Short Term Memory (LSTM) Networks

In a heavy-metal-ferromagnetic-metal-hetero-structure-based spintronic device, which exhibits perpendicular magnetic anisotropy (PMA), an experimental demonstration is presented here of multiple synaptic states being modulated electrically to obtain long-term potentiation (LTP) and long-term depression (LTD). Next, on-chip learning is simulated in a crossbar-array-based Long Short Term Memory (LSTM) network, where the model of each synaptic element in the crossbar array is based on the experimentally obtained LTP and LTD characteristics of the spintronic device and hence includes experimentally measured non-linearity, asymmetry, and cycle-to-cycle variations of LTP and LTD. Using the Airline Passenger time-series data set, we show that the LSTM can be trained successfully even in the presence of these non-idealities.