BR

BR(input_size, hidden_size;
    return_state = false, kwargs...)

Bistable recurrent network (Vecoven et al., 2021). See BRCell for a layer that processes a single sequence.

Arguments

• input_size => hidden_size: input and inner dimension of the layer.

Keyword arguments

• init_kernel: initializer for the input to hidden weights. Default is glorot_uniform.
• init_recurrent_kernel: initializer for the hidden to hidden weights. Default is glorot_uniform.
• bias: include a bias or not. Default is true
• recurrent_bias: include recurrent to recurrent bias or not. Default is true.
• independent_recurrence: hard-coded to true in this architecture. For the architecture without independent recurrence plese refer to NBR
• integration_mode: determines how the input and hidden projections are combined. The options are :addition and :multiplicative_integration. Defaults to :addition.

Equations

\[\begin{aligned} \mathbf{a}(t) &= 1 + \tanh\left( \mathbf{W}^{a}_{ih} \mathbf{x}(t) + \mathbf{w}^{a} \circ \mathbf{h}(t-1) + \mathbf{b}^{a} \right), \\ \mathbf{c}(t) &= \sigma\left( \mathbf{W}^{c}_{ih} \mathbf{x}(t) + \mathbf{w}^{c} \circ \mathbf{h}(t-1) + \mathbf{b}^{c} \right), \\ \mathbf{h}(t) &= \mathbf{c}(t) \circ \mathbf{h}(t-1) + \left(1 - \mathbf{c}(t)\right) \circ \tanh\left( \mathbf{W}^{h}_{ih} \mathbf{x}(t) + \mathbf{a}(t) \circ \mathbf{h}(t-1) + \mathbf{b}^{h} \right), \end{aligned}\]

Forward

br(inp, state)
br(inp)

Arguments

• inp: The input to the br. It should be a vector of size input_size x len or a matrix of size input_size x len x batch_size.
• state: The hidden state of the BR. If given, it is a vector of size hidden_size or a matrix of size hidden_size x batch_size. If not provided, it is assumed to be a vector of zeros, initialized by Flux.initialstates.

Returns

• New hidden states new_states as an array of size hidden_size x len x batch_size. When return_state = true it returns a tuple of the hidden stats new_states and the last state of the iteration.