In the "modern era", such model-parallel training techniques trace their roots back to AlexNet [14] and early influential systems such as DistBelief [6] and Project Adam [3].

In doing so, we are able to find the exact settings for which different methodsprovablybecome identical.

The resolution of observations is significantly larger in the ViZDoom environment, and there are more possible actions. In the ViZDoom environment, PLANS has access to more demonstrations to infer the underlying program. For additional information about the datasets, we refer to Sun et al. (2018). Figures 4 and 5 give examples of programs and demonstrations for the Karel and ViZdoom benchmarks respectively. In C.2, we detail which heuristics were used to improve the program and sequenceaccuracymetric. Results are reported in Table 5.

Machine learning models are famously vulnerable to adversarial attacks: small ad-hoc perturbations of the data that can catastrophically alter the model predictions.