This repository contains the code for "Training cascaded networks for speeded decisions using a temporal-difference loss" Michael L. Iuzzolino, Michael C. Mozer, and Samy Bengio [preprint]
All of the commands below are run from the parent google_research directory.
To train the model, launch the run.sh script. Specify base configuration parameters in configs/train_base_flags.py, which may be overwritten in run.sh flags.