New Publication: Making Models Disagree to Learn How Brains Compute

A new review from Tal Golan, Heiko H. Schütt, and Nikolaus Kriegeskorte, “Making models disagree to learn how brains compute,” has been published in Nature Reviews Neuroscience.

Neural network models provide a powerful way to express hypotheses about how the brain processes information, but their flexibility can make very different models appear similarly consistent with experimental data. The review explores an emerging solution: designing or selecting controversial stimuli that cause competing models to make different predictions.

By testing models where they disagree most strongly, researchers can design more informative experiments and more effectively distinguish between alternative theories of brain computation. The review brings together recent work on stimulus optimization and connects these approaches to Bayesian optimal experimental design.