Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session A: Tuesday, August 12, 1:30 – 4:30 pm, de Brug & E‑Hall

Universally Controversial Stimuli Reveal that Adversarial Robustness Improves DNN Prediction Accuracy across the Entire Human Auditory Cortex

David Skrill¹, Jenelle Feather², Samuel Victor Norman-Haignere³; ¹University of Rochester, ²Flatiron Institute, ³University of Rochester Medical Center

Presenter: David Skrill

Model comparison is central to all scientific progress. In sensory neuroscience, a key challenge is that distinct models often make similar neural predictions due to correlations between distinct features in the tested stimulus set. Here, we show how to distinguish models for a full neural population by designing a targeted “universally controversial” stimulus set that makes distinct, high variance predictions across an entire sensory cortical system (human auditory cortex) in every subject tested. We applied to compare the neural prediction accuracy of standard artificial neural networks (ANNs) from ANNs trained to be robust to “adversarial attacks”. Standard ANNs are notoriously vulnerable to small stimulus perturbations that can substantially alter the network’s decisions without meaningfully altering human perception. Yet, we find that the prediction accuracy of standard and robust ANNs in the human auditory cortex is virtually indistinguishable when measured using fMRI responses to natural sounds. In contrast, when tested with controversial stimuli, the cortical prediction accuracy of the robust model remains high throughout the auditory cortex, while the predictive power of the non-robust model drops to near zero. Universal controversiality thus opens the door to much more powerful model comparisons in sensory neuroscience and demonstrates a strikingly uncontroversial model improvement from adversarial training.

Topic Area: Visual Processing & Computational Vision

Extended Abstract: Full Text PDF