Bioinspired Fully On‐Chip Learning Implemented on Memristive Neural Networks

Inspired by biological neural systems, recent advances in neuromorphic computing have paved the way for intelligent devices. However, traditional memristive neural networks (MNNs) are incapable of independently determining high‐precision target weight and bias updates without relying on off‐chip intelligent machines, rendering the completion of the training process entirely on‐chip unattainable. Consequently, the critical challenge of achieving fully on‐chip learning remains to be addressed urgently. To address this bottleneck, this study introduces a novel MNN paradigm that facilitates fully on‐chip learning. Critically, this neuromorphic system adopts an operational mechanism derived from biological principles—contrastive Hebbian learning. This allows fully on‐chip learning using only voltage differences between pre and postsynaptic neurons and basic multiplications. Moreover, the experimental results demonstrate that the MNN achieves an accuracy (ACC) exceeding 98% in a pattern recognition task, and generalization error rates remain below 10% across all categories. This memristive neural system also exhibits ultralow power consumption, averaging 0.321 nJ per input image on a device, while maintaining over 80% accuracy within the conductance perturbations of ±1 μS, which reflects the adaptive restoring capability of these memristive neurons. Generally, this work advances efficient, bioinspired neuromorphic chips toward on‐chip learning and broader applications.