Figure 4

Generalization of the brain-inspired SFNNs to ANNs. (a) Fully-connected ID-net (Fig. 1b), but with sigmoid activation functions for the units (“Methods”), which is trained on a dataset that is similar to Fig. 2a (“Methods”). The blue and orange histograms, averaged over 720 instances, are constructed similarly to Fig. 2a for the three types of imperfect test sequences: fast/slow/wrong. The overlap between the blue and the orange histograms in the wrong panel is 0.56% for threshold 0.9 and zero for the other two panels. (b) Similar to Fig. 3, where the LeNet-5 architecture with ReLU activation function is trained on one individual’s handwriting, comprising a sequence of 10 different CIFAR-10 objects with 50 additional similar synthetic sequences (“Methods”). The panels represent the four imperfections: fast, slow, wrong, handwriting of different individuals, similar to Fig. 3a–d. The overlap between the orange and blue histograms vanishes for threshold 0.69, where each histogram was averaged over 720 instances comprising 20 samples (“Methods”).