iiinova — where your digital self keeps evolving

Mark R. Saddler

@msaddler.bsky.social

Postdoc at DTU Hearing Systems doing computational research in auditory neuroscience

Mark R. Saddler @msaddler.bsky.social · 8h

Interesting question! I unfortunately don't think we can say. Our DNN optimization isn't explicitly modeling evolution/development, but rather some opaque combination of the two that results in a similar end-state as the adult listeners we compared against. I suspect human thresholds are set by both

100

Mark R. Saddler @msaddler.bsky.social · 07/10/2026

Our preprint adds to the growing evidence that optimization for ecological tasks can explain human behavior beyond the tasks on which a model is trained, including a surprisingly comprehensive swath of classic psychoacoustic phenomena. [6/6]

110

Mark R. Saddler @msaddler.bsky.social · 07/10/2026

The model does not require internal noise to reproduce human perceptual limits. These results raise the possibility that absolute discrimination thresholds are largely determined by linear separability in representations optimized for natural behavior, rather than by intrinsic neural noise. [5/6]

removing stochasticity from the model has very little effect on absolute thresholds

110

Mark R. Saddler @msaddler.bsky.social · 07/10/2026

This match depended on whether the model was optimized for natural behavior. If the model was not optimized, its thresholds were worse than those of humans. And if the same architecture was optimized for individual psychophysical tasks, it exhibited superhuman sensitivity. [4/6]

model produces close-to-human thresholds, but only when optimized for ecological tasks

100

Mark R. Saddler @msaddler.bsky.social · 07/10/2026

Despite never being fit to human data or trained on psychophysical tasks, the model's thresholds closely matched human thresholds across a large battery of classic experiments spanning auditory sensitivity (1632 thresholds across 31 experiments!) [3/6]

table of 31 classic psychoacoustic experiments

100

Mark R. Saddler @msaddler.bsky.social · 07/10/2026

We trained a deep neural network to jointly localize and recognize words, voices, and environmental sounds from simulated auditory nerve input. We then measured the model's psychophysical thresholds using linear classifiers operating on the learned representations. [2/6]

schematic of linear classifier making psychoacoustic judgments using learned representations

100

Mark R. Saddler @msaddler.bsky.social · 07/10/2026

Pleased to share my new preprint with @joshhmcdermott.bsky.social and Torsten Dau: www.biorxiv.org/content/10.6... ! It shows that many of the characteristic limits of human hearing emerge from perceptual representations optimized for everyday hearing behavior. [1/6]

biorxiv.org

1114