Deconstructing Echo State Networks (ESNs)
Explore Reservoir Computing, the Echo State Property, closed-form PyTorch linear readouts, and forecasting chaotic systems 1000x faster than BPTT.
Part 1
The End of BPTT
The math behind Reservoir Computing, the Echo State Property, Spectral Radius, and why we can replace gradient descent with closed-form linear algebra.
Part 2
PyTorch Implementation
Building the massive, sparse, non-trainable random reservoir and Ridge Regression (Tikhonov Regularization) readout layer in pure PyTorch.
View Code on GitHub
Part 3
Predicting Chaos
Benchmarking our ESN against a standard PyTorch LSTM on predicting the chaotic Mackey-Glass sequence to demonstrate massive training speedups.