Neurotransmitter-Mediated Plasticity in 2D Perovskite Memristor for Reinforcement Learning

The neuromorphic computing architecture is a promising artificial intelligence for implementing hierarchical processing, in-memory computing, event-driven operation and functional specialization in computing systems. However, current investigations mainly focus on unisensory processing without objective experience which is contrary to the flexible sensory learning capability in the human brain that can sense and process information according to the ever-changing environment. For example, a dominant paradigm for reconfigurable bio-learning features is the emotional experience. The neurotransmitter dopamine is released during arousal, influencing the vital brain functions involved in cognition, reward learning, movement and motivation. Here, the on-demand configuration of a biorealistic synaptic connection based on a 2D CaTa2O7 (CTO) device is demonstrated that can be adaptively reconfigured for a reinforcement learning purpose by the light-active resistive switching, which originated from the photon-regulated metaplasticity. The low energy consumption of 12.4 fJ endows the reinforcement learning system with high power efficiency and reliability. Finally, in-sensor computing with a CTO synapse is implemented with a filtering function to process digital data in a neuromorphic engineering manner. This work demonstrates the feasibility of 2D perovskite neuromorphic device with enhanced biological plausibility in the approaching post-Moore era. Multisensory learning system based on a single CaTa2O7 nanosheet memristor is demonstrated. Benefiting from the bioinspired microstructure design, the device achieves a reliable light-active resistive switching capability, along with the metaplasticity engineering features, which enables the higher-order cognitive functions with reinforcement learning for AI algorithms.image