Fig. 3: Shape recognition.
From: Multifunctional human visual pathway-replicated hardware based on 2D materials
a Workflow of shape recognition in the human visual system. The schematic diagram is adapted with permission from refs. 47,48. LGN lateral geniculate nucleus, V1 primary visual cortex, V2 secondary visual cortex, V4 extrastriate cortex, IT inferotemporal cortex, NN neural network. The hardware structure of sparse neural networks with an orientation selector. b Test photocurrent results (marked with the gray shading in the background) for the convolution of the 50 × 50 pixel regular hexadecagon light input with eight 5 × 5 orientation convolution kernels (OCKs). c Schematic of the hardware principle for shape recognition. The light-mask input is generated by illuminating the light mask consisting of a right-angled triangle and the frosted glass. The right-angled triangle input is divided into five regions for OCK convolution, which are processed through the 5 × 8 × 4 sparse neural network. Simulation (d) and experimental conductivity (e) weights of the double-layer sparse neural network with 30 epochs. f Simulation (red) and experiment (blue) results of the recognition rate with 30 epochs. g Comparison of this sparse neural network and the fully connected neural network in the recognition rate, loss, device usage, and programming energy for one operation. The error bars represent the standard deviation.
