Information coding strategies in biological reservoir computing for multiclass time-series classification

Abstract

Physical reservoir computing (RC) using biological neural networks (BNNs) offers a framework for probing how high-dimensional dynamics arising from multicellular systems support complex information processing in the brain. However, how electrical inputs and extracellular readouts should be designed to suit different task demands remains unclear. Here, using primary rat cortical neurons cultured on a high-density microelectrode array, we investigated how input encoding and spike detection methods influence neural representations and the resulting RC performance in BNNs. Comparison of pulse amplitude modulation and latency coding revealed that the latter exhibited superior performance in demanding multiclass conditions, including a 10% improvement in a 10-class spoken digit classification task. Spike sorting further improved classification performance by enhancing class separability, while largely preserving the low-dimensional structure of population activity. These findings highlight the importance of designing information coding strategies according to task difficulty.