Fig. 5: Perovskite nickelate networks for spoken digit recognition and seizure detection.

a, Processing pipeline for spoken digit recognition task. The Lyon’s ear model, inspired by the human cochlea, is used to convert the audio signals into analogue signals with 64 frequency channels, which were converted into binary voltage spike trains via thresholding to be fed into the spatiotemporal processing layer. At each time step, a 5 V voltage pulse with 500 ns width was applied whenever there was a spike. Spike trains processed by the preprocessing layer were applied to a linear output layer to classify the inputs into the correct digits. b, Evolution of the hydrogen cloud thickness from 128 Pd electrodes. The results are compared for cases where every two electrodes are isolated with only temporal characteristics (temporal only) versus when all 128 electrodes exhibit spatial interactions (spatial and temporal). c, Accuracy of speech recognition is plotted against the number of times (1, 2, 4) the spike train is sampled for no preprocessing, temporal-only and spatiotemporal cases. The results show that the spatiotemporal processing by protonic nickelate networks increases the digit recognition accuracy. d, Processing pipeline for early seizure detection. EEG signals recorded from 23 electrodes placed on the scalp according to the International 10–20 system. The displayed signals include segments of both normal brain activity and seizure episodes. A threshold is applied to convert the continuous signal into spike trains and applied into the preprocessing layer classified by a linear output layer as either seizure or normal. e, A representative seizure event lasting 10 s shown for 3 channels. Time points for early detection are marked as 1 s, 2 s and 3 s. f, Seizure detection accuracy results show that after the spatiotemporal processing by protonic nickelate networks offers a substation advantage for early detection of seizures with high accuracy.

Source data