Reconfigurable, Reliable and Online Learning Enabled Memristive Neuromorphic Core Design for Brain-Inspired Computing

For several years, the von Neumann architecture has served as the foundation of contemporary computers due to its straightforward, cost-effective architecture for processing units and memory. However, this design encounters a significant impediment in meeting the growing demand for increased parallelism in complex computations. Additionally, the post-Moore's law era emphasizes the need for energy-efficient computing with fewer resources and reduced space. As a response to these challenges, researchers are actively seeking alternatives to the von Neumann architecture, with neuromorphic computing emerging as a promising candidate. Neuromorphic computing garnered attention initially for its ability to harness the parallelism inherent in bio-inspired networks, particularly on a standalone chip. The demand for real-time performance further elevated the appeal of neuromorphic computing, prompting various research teams to explore hardware implementations of neural networks. Simultaneously, advancements in neuroscience contribute to a deeper comprehension of biological neural systems, offering possibilities to bridge the disparity between the functions of biological neurons and those of artificial neural networks. Consequently, the concept of neuromorphic computing has gained increasing popularity. This work presents a reconfigurable mixed-signal neuromorphic core using memristors for improved efficiency. Memristors are nano-scale devices that can provide area and power efficiency for the overall system design and can be used to design synapses to regulate the weighted current accumulation into neurons. This research uses the test results from memristive devices composed of hafnium oxide to validate the synaptic design choices. It also optimizes its performance for process variation. This neuromorphic system allows runtime adaptation using programmable bio-inspired features implemented in the neuron. Several online learning methodologies have been implemented for this hardware system. Among them, the performance of the most popular method for online learning, namely spike-timing-dependent plasticity (STDP) has been evaluated for this system using a software model.