Adaptive Computation Time

tags
Neural networks, Algorithm

Adaptive computation time (ACT) was introduced in (Graves 2017) as a way to make computations in RNN adaptive. The network learns how many computational steps to use before emitting an output.

This is done by outputting an extra halting probability at each update step, and considering two timelines:

  • the input timeline which plays the role of an outer loop, at each of those step, a new input symbol is fed to the RNN. This step outputs a single output vector.
  • the internal processing timeline, this is the inner loop being run at each of the input steps. This runs until the cumulative halting probability is above a threshold and emits as many output values as steps..

Bibliography

Graves, Alex. 2017. “Adaptive Computation Time for Recurrent Neural Networks.” arXiv:1603.08983 [Cs], February.


← Back to Notes