Design of Robust Memristor-Based Neuromorphic Circuits and Systems with Online Learning

Computing systems that are capable of performing human-like cognitive tasks have been an area of active research in the recent past. However, due to the bottleneck faced by the traditionally adopted von Neumann computing architecture, bio-inspired neural network style computing paradigm has seen a spike in research interest. Physical implementations of this paradigm of computing are known as neuromorphic systems. In the recent years, in the domain of neuromorphic systems, memristor based neuromorphic systems have gained increased attention from the research community due to the advantages offered by memristors such as their nanoscale size, nonvolatile nature and power efficient programming capability. However, these devices also suffer from a variety of non-ideal behaviors such as switching speed and threshold asymmetry, limited resolution and endurance that can have a detrimental impact on the operation of the systems employing these devices. This work aims to develop device-aware circuits that are robust in the face of such non-ideal properties. A bi-memristor synapse is first presented whose spike-timing-dependent plasticity (STDP) behavior can be precisely controlled on-chip and hence is shown to be robust. Later, a mixed-mode neuron is introduced that is amenable for use in conjunction with a range of memristors without needing to custom design it. These circuits are then used together to construct a memristive crossbar based system with supervised STDP learning to perform a pattern recognition application. The learning in the crossbar system is shown to be robust to the device-level issues owing to the robustness of the proposed circuits. Lastly, the proposed circuits are applied to build a liquid state machine based reservoir computing system. The reservoir used here is a spiking recurrent neural network generated using an evolutionary optimization algorithm and the readout layer is built with the crossbar system presented earlier, with STDP based online learning. A generalized framework for the hardware implementation of this system is proposed and it is shown that this liquid state machine is robust against device-level switching issues that would have otherwise impacted learning in the readout layer. Thereby, it is demonstrated that the proposed circuits along with their learning techniques can be used to build robust memristor-based neuromorphic systems with online learning.