Post-initial submission follow-up analyses of model/human behavior

[judithfan]

• Extract those trials where humans: (a) consistently succeed; (b) are close to chance; (c) systematically fail.
  These can get passed into curiophysics for interestingness annotation.
• More detailed error analysis: on which scenarios / instances did humans and models diverge the most?
  Extract those trials where Visualize some model predictions -- can we tell why some of the vision models fail/succeed some trials?
• Which models’ behavior were most similar to which other models’?