COMPARATIVE ANALYSIS OF ARTIFICIAL, SPIKING, AND BIOLOGICAL NEURAL NETWORKS FOR RESERVOIR COMPUTING

Understanding the principles behind the superior learning and energy efficiency in the brain is critical for addressing fundamental limitations in current machine learning. Recently, physical reservoir computing with in vitro biological neural networks (BNNs) have gained significant attention to constructively explore the computational potential of BNNs. However, whether and how in vitro BNNs possess an computational advantage over artificial neural networks (ANNs), a dominant paradigm in current machine learning based on firing rate representation, and spiking neural networks (SNNs), a more biologically plausible form of neural networks simulating membrane potential fluctuations and signal transmission through action potentials, has remained unclear due to lack of suitable experimental platforms. In this study, we developed an all-optical closed-loop system based on calcium imaging and optogenetics to implement a two-class image recognition task in cultured rat cortical neurons. MNIST digits were illuminated directly to the neurons transfected with channelrhodopsin as optical patterns. Output was then generated from the evoked activity with a linear decoder, training the output weights by FORCE learning. We found that, at an equivalent network size, BNNs achieve classification accuracy equal to or higher than ANNs and SNNs in this task. We interpret this biological advantage as arising from subcellular morphological features, such as axon and dendrite, playing a critical role in enriching information representation in BNNs. The study is funded by MEXT Grant-in-Aid for Transformative Research Areas (A) “Multicellular Neurobiocomputing,” JSPS KAKENHI, the WISE Program for AI Electronics by Tohoku University, and RIEC Cooperative Research Projects at Tohoku University.