The geometry of telling things apart
Monkeys and human volunteers performing two distinct perceptual tasks made less accurate decisions when they did not know which task was coming, while an artificial neural network trained on the same tasks showed no such drop. Comparing a network trained to produce correct choices with one trained to reproduce the participants' choices pointed to feature interference, and under uncertain conditions the neuronal measurements showed stronger representations of irrelevant features and entangled representations of different features.[1]
In juvenile and adult zebrafish trained on an odour discrimination task, the same question was approached from the other side. Odour discrimination training selectively enhanced the separation of manifolds representing task-relevant odours from other representations, and manifold capacity predicted odour discrimination across individuals. The measurements came from telencephalic area pDp, the homologue of piriform cortex, and no obvious signature of attractor dynamics appeared.[2]
The two experiments turn on one variable. In both, the accuracy of a decision follows how far apart the representations of the relevant features are held: training pulls task-relevant manifolds apart and discrimination improves, while uncertainty entangles feature representations and accuracy falls.[1], [2]
The denominator
This is a mechanism, and it deserves to be taken seriously. It also deserves its denominator. The two experiments do not measure a clinical effect; one uses juvenile and adult zebrafish, the other monkeys and human volunteers in perceptual tasks. In both, the endpoint is accuracy on a laboratory task.[1], [2]
The strongest reading of the interference result is that a limit which looks cognitive has a measurable physical form: representations of irrelevant features grow stronger and crowd out the relevant ones. A weaker reading is available too. The measures differ between the two experiments, manifold capacity in one and feature interference in the other, and they may describe two separate phenomena instead of one shared constraint. The authors offer the capacity claim as a suggestion and do not present it as a demonstrated result.[1], [2]
The measurement that has not been made
The missing step is the one between the geometry and a person. A useful next measurement is narrow and specific: whether the separability of representations, measured the way these experiments measure it, tracks performance in someone whose attention is genuinely impaired, and whether it moves when the impairment moves. Until an experiment of that shape exists, the geometry stays limited to explaining the laboratory task it was measured on.[1], [2]