Memristive Neural Networks-Assisted Training Algorithms: Current State and Future Challenges

Existing challenges in modern AI architectures, including the von Neumann bottleneck and high energy demand, have led to increased research efforts towards the development of novel computing paradigms based on memristive technologies. This comprehensive review summarizes algorithmic approaches for training memristive neural networks (MNNs), including Memristive Spiking Neural Networks (MSNNs), integrating solutions implemented on hardware platforms with those developed in software. In contrast to existing reviews that consider algorithmic aspects within the broader context of hardware implementation, this review provides an algorithm-centric perspective, starting from training algorithms and analyzing different approaches to their adaptation to memristive hardware. The survey covers traditional machine learning algorithms such as backpropagation and its enhancements toward hardware-friendly solutions, as well as gradient-free methods, with particular emphasis on algorithms that address device non-idealities. We additionally provide an overview of bio-inspired learning rules, including Spike-Timing Dependent Plasticity (STDP) and reward-modulated STDP, which can be mimicked using memristive devices, along with reinforcement learning algorithms. Special attention is given to biologically plausible models of spiking neuron-astrocyte networks, and their direct implementation in memristive hardware, which significantly increase the efficiency of self-learning algorithms. Finally, this review presents an algorithm-centric perspective on open questions and challenges toward the large-scale implementation of memristive devices in AI systems.