In this paper we propose a generalization of this work that generally exhibits improved performace, but from an implementation point of view is actually simpler.

Neural Information Processing Systems http://nips.cc/

Error bound conditions (EBC) are properties that characterize the growth of an objectivefunction when apoint ismovedawayfrom theoptimal set.

We consider deep policy learning with only batched historical trajectories. The main challenge of this problem is that the learner no longer has a simulator or "environment oracle" as in most reinforcement learning settings.

Our framework is the infinite-horizon discounted Markov Decision Process (MDP).

Yet, same as the classical setup, the goal is still to compete against the best context arm at each round.

We agree with the reviewers that reporting classification accuracy is important. As described in the paper and as noted by Reviewer 1, augmentation by adding more26 channels has been used inthe contextofResNets (e.g.

Acommon formulation of curiosity-driven exploration uses the difference between the real future and the future predicted by a learned model [1].

Deepneural Arguably (open-sourced) evaluate time-consuming significance ership security ofpoison-only explore behaviors andpro watermark tousethe tion.

The first challenge is the exorbitant cost of scaling up the size of manually-labeled video datasets. The recent creation of large-scale action recognition datasets [5,15,25,26]hasundoubtedly enabled amajor leap forwardinvideo models accuracies.