Fig. 3: Schematic of the working principle of spectral reconstruction.
From: Sub-picojoule-per-bit volitional neuromorphic devices for precise targeting and tracking
a Parameter input. By employing nonlinear transformation expansion, the dimensionality of the input space is increased to 16 dimensions. b Residual neural networks (ResNet) training. c Spectral reconstruction of white light. d Average reconstruction accuracy for all voltage combinations under 450–800 nm. e The corresponding voltage combination for 733 nm light signals. The optimal voltage combinations marked as red squares are assigned to the minimum value of Idiff-pc. f The accuracy change curve for the three primary colors (460 nm, 554 nm, and 697 nm) and the selected 733 nm. The shadings represent the corresponding error bars for each color, and the “star” symbols suggest the optimal gate voltage for achieving maximum spectral reconstruction accuracy.
