Content

In this supplementary material folder, you'll find 3 folders with additional results (one folder per game).

Game folder

In each game folder, there is :

• A RISE folder with Gifs & mp4 videos of RISE saliency maps when perturbing different frames for a single game (the same game for all).
• An Occlusion folder with Gifs & mp4 videos of Occlusion saliency maps when perturbing different frames for a single game (the same game for all).
• 10 gifs of gameplay with different seeds
• The RISE, Occlusion screenshots that appear in the paper.
• 4 computational graphs examples (a single prediction per graph) in png format. One of them appears in the paper (appendix).
• A folder called "seed1" that has ALL the computational graphs for inference for all steps on a game (wiht seed 1).

Additional info for the seed1 folder

If a game took, for example, 1000 steps (saw 1000 new frames), the folder will have 1000 computational graphs files. Each file is the inference graph of how the program reacted to the new inputs, and led to the outputs/actions it gave when playing. The svg files load images that are in folders in the "seed1" folder. When viewing a single computational graph, for example the one named "p239.svg", one has to first run the command ln -s ../../../sup_mat sup_mat inside the "seed1" folder, so that the svg can find the images that it has to load.

We only need to make one symbolic link. The symbolic link is then used by all svgs to find the path to the images it has to load.

For converting the svg files to png we used a command like this : inkscape --export-filename=s1_503.png seed1/p503.svg.

Code

Code to run experiments and to visualize/calculate saliency maps is also included.

PONG

S1	S2	S3	S4	S5

S6	S7	S8	S9	S10

Freeway

S1	S2	S3	S4	S5

S6	S7	S8	S9	S10

Bowling

S1	S2	S3	S4	S5

S6	S7	S8	S9	S10

RISE perturbations

Pong	Freeway	Bowling

Occlusion sensitivity perturbations

Pong	Freeway	Bowling