,
9 tweets,
4 min read
Read on Twitter
Very cool paper, @sashankpisupati @anne_churchland! Does not look like what we get in instructed adult human subjects, though (e.g., doi.org/10.1016/j.neur… or doi.org/10.1101/439885). Another illustration of the description-experience gap, here across mice and men? [1/9]
In doi.org/10.1016/j.neur…, using a similar categorization task in humans, fitted lapse rates are only about 1-2% and typically lose against lapse-free accounts. And rightly so, if subjects have understood that task rules *never* change (which mice haven’t). [2/9]
In doi.org/10.1101/439885, using a bandit task where exploration now has a clear adaptive value, fitted human lapses are also very rare (5% on average). Instead, non-greedy/suboptimal decisions are value-dependent, driven by learning noise and a softmax choice policy. [3/9]
More strikingly, we compared conditions where exploration is either adaptive or useless (by hiding or displaying foregone rewards). Despite this difference, which modulates the softmax temperature, fitted lapse rates are very similar in the two conditions (6.4% vs. 5.6%). [4/9]
The fraction of non-greedy decisions was modulated by local surprise, as in Thompson sampling (and in mice), but the effect in humans could be fully explained by the multiplicative structure of internal noise, rather than by active uncertainty-guided exploration. [5/9]
It looks like humans, trained partly by verbal/written description, are not continuously learning and updating their understanding of task rules. Their lapse rates are minimal (reflecting motor errors and inattention then?) and unaffected by uncertainty about task rules. [6/9]
By contrast, animals are trained by experience over days/weeks/months and seem to be continuously monitoring task rules for possible changes. This is very important, because it says that most lapses in mice/animals do not reflect task disengagement nor motor errors. [7/9]
It reminds me of recent work (doi.org/10.1101/489450) by @kishoreneuro et al. about latent knowledge unexpressed behaviorally by animals during learning of task rules. Tested animals, incl. mice, are learning faster and better than standard behavioral metrics suggest. [8/9]
However, I would not at this stage extrapolate what was found in mice to the human lapses observed in instructed experiments. Of course, one could make human subjects learn task rules by experience (as in mice), and see what happens to their lapse rates… [9/9]
