Making models disagree to learn how brains compute.

Optimizing controversial stimulus sets can make neural network models disagree on predictions, improving discrimination among brain information-processing hypotheses despite high model capacity.

• Why it matters: Standard comparisons using stimuli from the training distribution often fail to reveal differences between models because they can approximate arbitrary functions.
• What they did: The Review describes methods that optimize stimulus sets for model discrimination and use out-of-distribution generalization, selecting or synthesizing stimuli to maximize model-comparison power.
• The result: By tempering stimulus choice with a prior, controversial stimuli can combine naturalistic ecological validity with artificial-stimulus discrimination power, linking the approach to Bayesian optimal experimental design.

AI summary of the abstract · may contain errors · not medical advice · read the paper.