English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

We study learning of recurrent neural networks that produce temporalsequences consisting of the concatenation of re-usable "motifs". In the contextof neuroscience or robotics, these motifs would be the motor primitives fromwhich complex behavior is generated. Given a known set of motifs, can a newmotif be learned without affecting the performance of the known set and thenused in new sequences without first explicitly learning every possibletransition? Two requirements enable this: (i) parameter updates while learninga new motif do not interfere with the parameters used for the previouslyacquired ones; and (ii) a new motif can be robustly generated when startingfrom the network state reached at the end of any of the other motifs, even ifthat state was not present during training. We meet the first requirement byinvestigating artificial neural networks (ANNs) with specific architectures,and attempt to meet the second by training them to generate motifs from randominitial states. We find that learning of single motifs succeeds but thatsequence generation is not robust: transition failures are observed. We thencompare these results with a model whose architecture andanalytically-tractable dynamics are inspired by the motor thalamocorticalcircuit, and that includes a specific module used to implement motiftransitions. The synaptic weights of this model can be adjusted withoutrequiring stochastic gradient descent (SGD) on the simulated network outputs,and we have asymptotic guarantees that transitions will not fail. Indeed, insimulations, we achieve single-motif accuracy on par with the previouslystudied ANNs and have improved sequencing robustness with no transitionfailures. Finally, we show that insights obtained by studying the transitionsubnetwork of this model can also improve the robustness of transitioning inthe traditional ANNs previously studied.

Thalamocortical motor circuit insights for more robust hierarchical control of complex sequences