What values are saved between episodes?

[olegkolner]

Hello everyone!

I read an interesting issue about memory module #19 and as I have understood, the controller's (LSTM Cell) memory is cleaned after each episode (i.e. after a sequence is processed). Then my question is: what values, which have to be learned, will not be cleaned and, hence, passed over between episodes?

[dm-jrae]

The trainable model parameters! I.e. the trainable weights and biases in the controller, dnc core and addressing module. If you want to actually preserve the state of the dnc between episodes, you can just pass the state along between episode boundaries, instead of calling initial_state for a fresh state, or use tf.contrib.rnn.static_state_saving_rnn to facilitate the saving and loading of states.

[olegkolner]

Thanks for the rapid response!