Neural networks predicted the shortcut mice would struggle to unlearn

A preliminary lesson can make a later problem easier while also encouraging a stubborn mistake. University of Utah researchers explored that possibility by training recurrent neural networks, then testing their predictions in mice. Their September study follows a specific error from the artificial system into behaviour and recordings from a living brain.[2]

The tempting shortcut

Mice encountered pairs of odour pulses lasting two or five seconds. A water reward followed a correctly timed lick after unequal pairs; short–short pairs required withholding a lick. Practising only short–long pairs encouraged licking after a long pulse. When long–short trials arrived, that response came too early, before the second pulse had finished.[1]

Four mice received this restricted practice, compared with 16 given both unequal pair types. After eight days on the full task, the restricted group still performed significantly worse. Their recurring error matched the network prediction. Because short–short trials made up half the task, the shortcut could theoretically achieve 75% correct while failing every long–short trial.[1]

This illustrates why a sequence of successful responses can leave the wrong rule intact. A learner may encounter enough confirming examples to keep using its shortcut. Including a second kind of example can change what the early task demands, even before the later task becomes more difficult.[1]

A prediction inside the brain

The researchers also compared the changing patterns of activity. Utah's account describes a cycle through which both appropriately trained networks and mice progressed during a trial, with a separate deviation around the moment a response was required. That organisation helps distinguish keeping track of the unfolding trial from deciding when to act. Poorly structured training produced different activity and behaviour.[2]

The brain region was selected for a reason. In a separate 2024 study, Erin Bigus and colleagues examined the medial entorhinal cortex, where cells can fire at particular moments during an interval. Chemically suppressing activity there interfered with learning context-dependent timing. That earlier intervention supports the region's relevance; the new computational experiment asks how experience might organise the activity it carries.[3]

The networks were simplified task models, not biological reconstructions, and the experiment did not test human pupils. Their value was a checkable prediction.[1]

There is a wider theoretical reason to study such practice sequences. An ICML 2024 paper by Jin Hwa Lee, Stefano Sarao Mannelli and Andrew Saxe analysed how learning component skills can affect learning a compound task. Its mathematical models found that the usefulness of preliminary training depends on task complexity and the strategy used. The important design choice is which pieces of a problem early experience makes available.[4]

Jack Bowler told Utah's communications team that modelling first helped generate hypotheses faster and reduce the animals needed for testing. That gives AI a practical role: make a learning mechanism specific enough for an experiment to challenge it.[2]

Report an error or suggest a correction