Extended Data Fig. 4: Sequence prediction details.
From: Deep learning incorporating biologically inspired neural dynamics and in-memory computing
The values in all the panes of this figure were obtained by averaging over ten different initializations. Standard deviation is reported along the results and marked with error bars in the plots. a, Language modelling training perplexity evolution for SNU- and sSNU-based architectures. b, Comparison of test perplexity with other results70,81,89. ANN results using standard architectures with similar training techniques were considered, that is no pre- or post-processing, single network, truncated BPTT, no dropout. WT denotes weight tying of the output layer with the embedding layer. c, Music prediction loss evolution for sSNU-based network. d, Comparison with other results71,72.
