CAN BIOLOGICAL NEURONS BE USED FOR MACHINE LEARNING? A HYBRID BIOLOGICAL–ARTIFICIAL NEURAL NETWORK CLASSIFIER

Artificial neural networks (ANNs) excel at solving complex problems across various domains, but biological neural networks (BNNs) surpass them in energy efficiency and parallel processing, sparking interest in biological computation. In bio-computation, BNNs are cultured in vitro, stimulated, and recorded predominantly via microelectrode arrays (MEAs). Recent approaches like "organoid intelligence" aim to train BNNs on specific tasks using biological learning paradigms. However, such systems often overlook fundamental limitations or risk overinterpreting the underlying electrical signals.
Our goal is to engineer a BNN with a defined architecture and controlled stimulation to investigate task-solving behavior in interpretable, well-defined conditions. Human induced pluripotent stem cell-derived NGN2 neurons are seeded on high-density MEAs, with neuronal growth spatially confined and directed by polydimethylsiloxane microstructures. This setup enables precise investigation of stimulation-induced network activity. We observe nonlinear neuronal responses, such as changes in firing rates or latency, depending on stimulation intensity or frequency, analogous to activation functions in ANNs. Utilizing this insight, we construct a feed-forward hybrid network in which BNNs are artificially interconnected. The stimulation-induced scalar response measures are linearly combined to define the input for downstream BNNs. The non-differentiable nature of biological neurons is accommodated through tandem learning. A multistage training strategy gradually transitions the model towards increasingly realistic conditions.
Using this framework, we create a hybrid network with a single hidden layer of 16 nodes that solves the Yin-Yang dataset with 86% accuracy. This illustrates an initial step toward practical supervised biological computation.