Fig. 1: Local circuit model of a cANN to reconstruct cerebellar language functions.
From: Emergence of syntax and word prediction in an artificial neural circuit of the cerebellum
a Previously proposed block diagram with Inputs/outputs for next-word prediction. b The three-layer cerebellar circuit, consisting of a feedforward pathway (brown) and a recurrent pathway (blue). The climbing fiber pathway (gray arrows) delivers the prediction error signal to update the synaptic weights on the Purkinje and output cells (gray shading). c, d Learning curves for the prediction error (c; cross entropy; median and IQR) and percentage correct predictions (d) for 20 cANNs trained on next-word prediction. Inset, magnified view of the early learning period (gray interval). e Predictions (five candidate words) for the two instances of “the” in a sentence by a well-trained cANN. Correct predictions are highlighted in red, and synonyms are in orange.
