[Re] [Ablation] Unsupervised State Representation Learning in Atari

Gabriel Alacchi, Guillaume Lam, Carl Perreult-Lafleur

Abstract:

In this study, we performed some ablations on the main model developed in the paper "Unsupervised
Representation Learning in Atari" [2] as part of the 2019 NeurIPS Reproducibility Challenge. In this
paper, Anand et. al introduce a new learning method called SpatioTemporal DeepInfoMax (STDIM),
which is an unsupervised method that aims at learning state representations by maximizing particular
forms of mutual information between a series of observations. Our work focuses on recreating a subset
of their results, along with hyperparameter tuning, slightly altering the STDIM learning objective, and
altering the receptive field of the encoder model that Anand et. al introduce in their article. We also
suggest directions for further expanding the STDIM method. Our results also suggest that creating an
ensemble model would allow for further boosting of the effectiveness of this model.

Track: Ablation

NeurIPS Paper Id: https://openreview.net/forum?id=HkelAVBeIr