Traditional hardware systems struggle with implementing current artificial neural networks due to the waste of substantial computational resources on insignificant data. Hardware realization of sparse neural networks offers a significant solution because of their potential to concentrate solely on crucial data. However, these devices still face great challenges in signal encoding and attention-guided sparse capture. Herein, we demonstrate a large-scale sparse-capture neural network (SCNN) using vertical multichannel photoelectrochemical transistors, which are constructed from the ultrashort, tri-layer, oxygen-gradient-engineered indium-tin oxide channel with an approximately 15 nm thick. This device exhibits high sparsity at a low operating voltage of 3.0 V, facilitating dynamic neural connectivity and outstanding energy efficiency. The proposed SCNN achieves recognition accuracy exceeding 94% and reduces energy consumption by over 30%. Therefore, this work offers a promising avenue toward energy-efficient neuromorphic systems for edge AI, real-time sensing, and adaptive decision-making.
Yu et al. (Fri,) studied this question.