Read on Twitter
Great study by @jwdegee @donner_lab out in @eLife on the effect of pupil-linked arousal on choice biases during stimulus detection and memory retrieval. They show how phasic pupil dilation predicts less biased choices, in humans and mice! My two cents below: (1/9) https://twitter.com/donner_lab/status/1273286311202500608
Their main finding is very consistent with our observations during reward-guided learning, where phasic pupil dilation predicts more switches - i.e., less biased choices in a bandit task where subjects repeat their previous choice on ~80% of trials. (2/9) https://doi.org/10.1038/s41593-019-0518-9
We also found that phasic pupil dilation predicts more variable learning. In other words, phasic pupil dilation seems to shift the ‘bias-variance’ trade-off of behavior - as very neatly shown by @alsfilip et al. in their recent paper, also in @eLife. (3/9) https://doi.org/10.7554/eLife.57872
Regarding the last tweet of the thread, our idea of computation ‘noise’ does include trial-to-trial variability in drift rate - even if it is driven by fluctuations in arousal (which may or may not be random) - as one of its possible sources. (4/9) https://doi.org/10.1016/j.neuron.2016.11.005
We don’t define noise as something genuinely random at all levels of description, still random if we had accounted for the activity of every neuron in the brain (and every neuromodulatory system). If we did define noise in this way, there may not be much noise left. (5/9)
Instead, we define noise in a relative sense, from the perspective of the input-output function used to solve the task. In other words, ‘effective’ noise that may not be random in an absolute sense, but still triggers variability across repetitions of the same task input. (6/9)
In this view, arousal-linked variability in drift rate is still noise - not as something ultimately unexplainable from the neuroscientist perspective, but as something that generates trial-to-trial variability from the perspective of the task to be solved. (7/9)
Although our definition of computation noise is more inclusive than the extreme definition, it is still different from output/policy noise (e.g., softmax) and from input/sensory noise. By different, I mean 1. different statistical signatures that can be fitted to behavior, (8/9)
and 2. different cognitive functions (see our new preprint with @FindlingCharles). We think the arousal-linked computation noise shown in this very insightful study by @jwdegee et al. may contribute to adaptive behavior in adverse conditions. (9/9) https://doi.org/10.1101/2020.06.10.145300
