Skip to content

francesco-mannella/Echo-State-Networks

Repository files navigation

Implementing Echo-State Networks in tensorflow2.

ESN-definition.ipynb Describes the mathematics of leaky echo-state networks together with an analytical method to correct the spectral radius of the inner weights accounting for leakiness and a matricial trick to improve the variance of network states during its dynamics.

ESN.py Contains the definition of a customized tensorflow Cell. The inizialization of the weights uses numpy function via tf.py_function because tf.self_adjoint_eigvals only works on self-adjoint matrices.

ESN-usage.ipynb Contains an example of training on a simple dataset.

parametric-sequence-learning.ipynb Contains a more complex example. A set of 2D trajectories is learned and generalization to the whole family of trajectories is tested.