Self-sensitizable neuromorphic device based on adaptive hydrogen gradient

Neuromorphic computing faces long-standing challenges in handling unknown situations beyond the preset boundaries, resulting in catastrophic information loss and model failure. These predicaments arise from the existing brain-inspired hardware's inability to grasp critical information across diverse inputs, often responding passively within unalterable boundaries. Here, we report self-sensitization in perovskite neurons based on an adaptive hydrogen gradient, transcending the conventional fixed response range to autonomously capture unrecognized information. The networks with self-sensitizable neurons work well under unknown environments by reshaping the information reception range and feature salience. It can address the information loss and achieve seamless transition, processing ∼250% more structural information than traditional networks in building detection. Furthermore, the self-sensitizable convolutional network can surpass model boundaries to tackle the data drift accompanying varying inputs, improving accuracy by ∼110% in vehicle classification. The self-sensitizable neuron enables networks to autonomously cope with unforeseen environments, opening new avenues for self-guided cognitive systems.