Fig. 7: Comprehensive performance evaluation of the baseline(green) and biomod (red) with optimized hyperparameters across diverse tasks.
From: Concept transfer of synaptic diversity from biological to artificial neural networks

a Classification error rates on four image datasets (CIFAR10, ImageWoof, CIFAR100, Tiny ImageNet) comparing state-of-the-art architectures (ResNext, WResNet, EfficientNet, SEResNeXt, Swintrans V2) with their biomod counterparts, showing consistent error reductions of 0.1–19.6%. b Time series prediction performance on Thomas and Lorenz96 benchmarks using various architectures (LSTM, GRU, FDN, Transformer), demonstrating NRMSE improvements of 3.4–11.2% through biological modifications. Small black whiskers denote standard deviations.