Fig. 4: Few-shot learning.
From: Neuromorphic overparameterisation and few-shot learning in multilayer physical neural networks
a System predictions (black lines) in the overparameterised regime when rapidly adapting to 50 training data-points (grey circles) at the beginning of the series. The PNN is able to rapidly learn the task and shows strong performance on the test data set (target points in red). Shading is the residual between true and predicted values. b Sine-transformations with sparse training data. The system is asked to learn five different frequency components of a time-varying function (dashed black line). Here, the PNN is shown just 15 sparsely separated training data points (grey circles). As the system is updated (coloured plots), the PNN prediction steadily improves. c Two example frequency components of the targets in b, which the PNN learns. d, e Average error,  <MSE>, and variance of the error,  <σ2MSE>, computed over 500 signal transformation tasks as the train length and updates vary. The system shows good generalisation across all tasks, as shown by the low variance. f Few-shot signal transformation using a single array after 500 updates. The single array fails completely.
