next up previous
Next: Limitations and Extensions Up: The Architecture and the Previous: On-Line Versus Off-Line Learning

Unfolding in time

An alternative of the method above would be to employ a method similar to the `unfolding in time'-algorithm for recurrent nets (e.g. [Rumelhart et al., 1986]). It is convenient to keep an activation stack for each unit in $S$. At each time step of an episode, some unit's new activation should be pushed onto its stack. $S$'s output units should have an additional stack for storing sums of error signals received over time. With both (4) and (5), at each time step we essentially propagate the error signals obtained at $S$'s output units down to the input units. The final weight change of $W_S$ is proportional to the sum of all contributions of all errors observed during one episode. The complete gradient for $S$ is computed at the end of each episode by successively popping off the stacks of error signals and activations analogously to the `unfolding in time'-algorithm for recurrent networks. A disadvantage of the method is that it is not local in space.

Juergen Schmidhuber 2003-02-13

Back to Recurrent Neural Networks page