where:
- h is the state of the leaky cell
- x is the input
- lambda is the leak rate
- Win is the input weight matrix
- Wh is the hidden weight matrix
- sigma is the nonlinearity or transfer function, typically, hyperbolic tangent
- a leaky cell module that can be used as a recurrent layer in any architecture
- a function to generate initial weights for the leaky cell
- a simple one layer leaky recurrent network trainable with backprop
- a reservoir implementation with a specific 'train' method to train the readout weights with a L2 regularized linear regression