Adversarial training greatly improves DNN predictions across human auditory cortex, preprint finds
GretaTuckute · x · 2026-10-10
A new bioRxiv preprint by David Skrill, Jenelle Feather, and Sam Norman-Haignere introduces a stimulus-synthesis method that decorrelates neural predictions from competing encoding models.
- The approach synthesizes auditory stimuli that isolate predictive differences between models, avoiding confounds from correlated predictions.
- Using this method, they show adversarial robustness training substantially improves DNN prediction accuracy across the entire human auditory cortex.
- The work offers a cleaner way to compare encoding models in sensory neuroscience and suggests adversarially robust models better match human neural representations.
More from AGI Musings
- Hillclimbing "predict the next user message" could be how we get RSI, researcher speculates — GabGarrett · 2026-10-10
- AI commentators: AGI disruption only goes mainstream when half of white-collar jobs vanish — iruletheworldmo · 2026-10-10
- DeepMind's Andrew Lampinen defends anthropomorphic metaphors for describing AI — AndrewLampinen · 2026-10-10
- DeepMind's Andrew Lampinen defends anthropomorphic metaphors for describing AI — AndrewLampinen · 2026-10-10
- Pure math and theoretical CS are quietly flooding into biology with good results — fkasummer · 2026-10-10
- 'Professional toe-treading': execs fear being bypassed by AI-armed colleagues, says industry insider — beglen · 2026-10-10