Handwritten digit recognition using a biologically-inspired physical reservoir of single-transistor chaotic oscillators

In this paper, we investigate handwritten-digit recognition using a compact physical reservoir computer based on a network of 59 single-transistor chaotic oscillators derived from a neuronal culture. Input images from the MNIST dataset are first reduced to 8 × 8 pixels and then encoded as node-dependent control voltages applied locally to the nodes of the oscillator network. A mapping strategy in which the most informative pixels, identified based on the variance or ANOVA F-value, are assigned preferentially to structurally central nodes of the reservoir network, is employed. These static images are transformed within the network into class-dependent spatiotemporal responses from which low-dimensional parameters, including temporal standard deviation, mean phase synchronization, and spectral measures, are extracted. The obtained features are subsequently processed by a shallow multilayer perceptron acting as the readout stage. Numerical simulations and physical experiments with an analog electronic setup show that the reservoir achieves its best performance in an intermediate coupling regime, where sufficient coordination coexists with dynamical heterogeneity. Under these conditions, the classification accuracy obtained from the numerical simulations reaches approximately 81%; significantly higher than the baseline obtained using only image pixels with the selected small-scale classification network. From experimental measurements on the circuit board, the best accuracy reaches 75% and is obtained when temporal variability parameters are combined with the raw image representation. Beyond classification performance, this study demonstrates that a biologically inspired topology can play a central role in small-scale physical reservoirs by determining how information is distributed, transformed, and separated through collective nonlinear dynamics. This study thus establishes chaotic electronic oscillator networks as viable and interpretable analog substrates for reservoir computing.