Fig. 4: Overview of our reservoir computing implementation with a 2D thermal neuristor array.
From: Collective dynamics and long-range order in thermal neuristor networks
The MNIST handwritten digit dataset44 is used as a benchmark. Each image from the dataset is translated into input voltages for a 28 × 28 thermal neuristor array. The array’s spiking dynamics are gathered as the reservoir output. A fully connected output layer, enhanced with softmax nonlinearity, is trained to classify the digit. The bottom-left panel illustrates the training process, displaying both loss and accuracy, culminating in a final test set accuracy of 96%. Source data are provided in the Source Data file.
