
Our new preprint is now available! Recurrent Spiking Neural Networks for Continuous Motor Control: A Biologically Plausible Model of Motor Cortex

Neuroscience. Tokyo.

Our new preprint is now available! Recurrent Spiking Neural Networks for Continuous Motor Control: A Biologically Plausible Model of Motor Cortex

Our brilliant graduate student Hiroto Imamura, now at UCSB, led this work from concept through all simulations. Check out his recent Nature Communications paper, too! nature.com/articles/s41467-0…

Several fundamental questions about this large-scale circuit (CC, BG, CB, Th) remain unresolved, but we believe the overall outlook is encouraging.

In the present study, we show that the motor cortex can actually learn motor control with its local synaptic plasticity. The teaching signals could be routed from basal ganglia and cerebellum.

In Yoshida et al. (J. Neurosci., 2025), we focused on the pathways from cortex to the basal ganglia and cerebellum. In Tanno et al. (bioRxiv, 2026), we turned to the return pathways from the basal ganglia and cerebellum, through the thalamus, back to cortex.

Interestingly, we also found that this local plasticity rule was consistent with experimental spine-plasticity data.

Kolen–Pollack mirroring learns feedback without copied weights or symmetric recurrence. Centered FSN learned Reacher from target actions and Ant from reward, echoing cerebellar and basal-ganglia teaching.

FSN returns errors as spikes, so each synapse learns from presynaptic activity × the feedback-evoked postsynaptic change.


We developed and combined three new methods—CBIT, SCI and FSN—to do both. CBIT learns from equilibrium firing rates, avoiding long spike simulations and BPTT.

Can an RSNN learn continuous movement with motor-cortex-like local plasticity? Existing models either control movement or resemble motor cortex, but not both. biorxiv.org/content/10.64898…

the cortex remains a passive recipient of what these subcortical structures have learned. In our next study, we will reveal how the motor cortex itself can learn.

This study was led by Sho Tanno, who is now a medical resident. We have now come full circle through the loops linking the cortex with the basal ganglia and cerebellum. At this stage, however,

layer, L1. Incorporating this mechanism further improved learning in the model. Together, this study identifies anatomical constraints and computational requirements for the feedback pathways returning to the cortex, strengthening the circuit basis of synergistic RL.

tasks requiring simultaneous regulation of two or more neuronal populations only when these functional modules were organized into fine-grained parallel loops. Another factor was dendritic spiking driven by basal ganglia and cerebellar feedback arriving in the same cortical

functional roles of individual cortical neurons was also important. We showed that this alignment could be established using information available locally within the cortex, through a process we call “local calibration.” Interestingly, the model could handle multivariable BMI