motorSRNN: A spiking recurrent neural network inspired by brain topology for the effective and efficient decoding of cortical spike trains.

• In this study, we proposed motorSRNN, a recurrent spiking neural network (SNN) that draws inspiration from the neural motor circuit of primates. • The motorSRNN advanced the performance of previously reported SNN method in similar CST-decoding tasks, a feedforward SNN (fSNN), by over 25%. • The energy efficiency of motorSRNN was theoretically approximately 1/50 compared to traditional GRU and LSTM architectures. • The motorSRNN outperformed fSNN, GRU, and LSTM in terms of early-classification capabilities from 2 ms to the end in the 50-ms sample duration. • The motorSRNN elucidates a plausible rationale for the biologically employed topology: to enhance the resilience against Poisson noise from adjacent neurons. Decoding firing rates, averaged from cortical spike trains (CST), has yielded significant progress in invasive brain-machine interfaces (BMI). CSTs are theoretically more informative and efficient than firing rates. By directly decoding CST, spiking neural networks (SNN) exhibit promise for enhancing invasive BMIs due to high compatibility with CST and a low-energy consuming nature. However, whether SNNs can decode CST with applicable performance in terms of classification accuracy and energy consumption remains unclear. In this study, we proposed motorSRNN, a recurrent SNN topologically inspired by the primate motor neural circuit. Employed to decode CST from the primary motor cortex of two monkeys performing 4-direction reaching tasks, the motorSRNN achieved average classification accuracies of 89.44 % and 79.87 % for the 4 directions, respectively. This outperformed previously reported SNN method in similar CST-decoding tasks, a feedforward SNN (fSNN), by more than 25 %. Furthermore, motorSRNN demonstrated superior early-classification capabilities compared to fSNN, GRU, and LSTM from 2 ms to the end in the 50-ms sample duration. Additionally, it only theoretically consumed around 1/50 energy compared to traditional GRU and LSTM. Finally, motorSRNN offers insights into a possible rationale for the biologically employed topology: to enhance the resilience against Poisson noise from neighboring neurons in the biological brains. In conclusion, our proposed motorSRNN is feasible for effective and efficient CST decoding, laying the preliminary groundwork for constructing a fully implanted neuromorphic BMI. [ABSTRACT FROM AUTHOR]