Q3 reuses the Q2-selected predictor and its train-only robust scaler. It does not train a different predictor just to make an attribution method look better.

The experiment compares Integrated Gradients with five-slot grouped occlusion on the held-out Attachment 2 validation split. It measures deletion comprehensiveness, sufficiency, local rank stability under small input noise, and runtime. Polarity and continuous intensity explanations are selected separately because the two outputs can depend on different evidence.

For the 20 Attachment 4 videos, the script reads the aligned feature pickle and matching MP4, predicts polarity/intensity, then applies the Q1 hard CTC Viterbi word-time procedure to each transcript and source audio. Text wordpieces and aligned audio/vision slots inherit the CTC word interval. The output includes CTC quality and validity flags; CTC timestamps are a weak temporal reference, not human event annotations or ground truth.

Run

The environment is the uv-managed Q2 environment, which contains the same CUDA PyTorch, Transformers, NumPy, and plotting dependencies:

cd deep_learning/Q3
uv run --project ../Q2 python -m q3.explain_selection

Results are written to deep_learning/Q3/outputs/explanation_selection/. Q2 model checkpoints and normalization statistics are read from deep_learning/Q2/outputs/algorithm_selection/.