Back
Subject-to-subject and session-to-session variability in behavioural strategy for the International Brain Lab task
Most models of rodent behavior fit responses averaged across sessions and animals. Here, we ask instead whether different mice use different strategies and whether the same mouse uses different strategies on different sessions. Our analysis is based on a large cohort of mice (N{\textequals}63), each performing 3-4 sessions of the International Brain Laboratory task. On each trial, mice earned liquid reward by turning a wheel to report whether a Gabor patch (of variable contrast) was on the left or right side of a screen. The prior probability that the stimulus appeared on the right switched in blocks between 0.2 or 0.8 in a random and uncued manner, with a nearly flat hazard rate. Here, we investigate different forms of variability in the behavior of the population using a 3 level hierarchical Bayesian model (HBM) which had a session level, nested within a subject level, itself nested within a population level. The session level captured the between-session variability within a subject; the subject level captured the between-subject variability. The HBM allowed L{\textequals}4 candidate categories of behavioral strategies to which each session of each mouse could be attributed. The first category of optimal Bayesian agents implement the full generative process of the task, and so enjoy an explicit notion of block structure. The second category of stimulus-kernel strategies captures a local form of the bias within a block by low-pass filtering the sides on which stimuli were recently presented. The third action-kernel category is similar to the second but integrates over the sides that the mouse recently chose rather than over stimulus sides. The final fixed belief category maintains a fixed bias that does not change throughout the session. Inference was performed by combining variational Expectation-Maximization for session- and mouse-level parameters with type II maximum likelihood for population-level parameters (Piray et al. 2019). We found substantial between- and within-subject variability. For the former: the model fits favored action-kernel strategies on at least one session for 71{\textpercent} of mice, optimal Bayesian strategies for 30{\textpercent}, fixed belief strategies for 19{\textpercent}, and stimulus kernel strategies for 13{\textpercent}. For within-subject variability, {\textasciitilde}50{\textpercent} of the mice that performed 2 sessions or more were better explained by multiple strategies across sessions. Our results highlight the importance of not averaging across animals and sessions when analyzing rodent behavior. We are currently analyzing an additional, within-session level of strategy switching, and considering the representation of these various systematic differences in neural activity.
@inproceedings{item_3353178, title = {{Subject-to-subject and session-to-session variability in behavioural strategy for the International Brain Lab task}}, booktitle = {{50th Annual Meeting of the Society for Neuroscience (Neuroscience 2021)}}, abstract = {Most models of rodent behavior fit responses averaged across sessions and animals. Here, we ask instead whether different mice use different strategies and whether the same mouse uses different strategies on different sessions. Our analysis is based on a large cohort of mice (N{\textequals}63), each performing 3-4 sessions of the International Brain Laboratory task. On each trial, mice earned liquid reward by turning a wheel to report whether a Gabor patch (of variable contrast) was on the left or right side of a screen. The prior probability that the stimulus appeared on the right switched in blocks between 0.2 or 0.8 in a random and uncued manner, with a nearly flat hazard rate. Here, we investigate different forms of variability in the behavior of the population using a 3 level hierarchical Bayesian model (HBM) which had a session level, nested within a subject level, itself nested within a population level. The session level captured the between-session variability within a subject; the subject level captured the between-subject variability. The HBM allowed L{\textequals}4 candidate categories of behavioral strategies to which each session of each mouse could be attributed. The first category of optimal Bayesian agents implement the full generative process of the task, and so enjoy an explicit notion of block structure. The second category of stimulus-kernel strategies captures a local form of the bias within a block by low-pass filtering the sides on which stimuli were recently presented. The third action-kernel category is similar to the second but integrates over the sides that the mouse recently chose rather than over stimulus sides. The final fixed belief category maintains a fixed bias that does not change throughout the session. Inference was performed by combining variational Expectation-Maximization for session- and mouse-level parameters with type II maximum likelihood for population-level parameters (Piray et al. 2019). We found substantial between- and within-subject variability. For the former: the model fits favored action-kernel strategies on at least one session for 71{\textpercent} of mice, optimal Bayesian strategies for 30{\textpercent}, fixed belief strategies for 19{\textpercent}, and stimulus kernel strategies for 13{\textpercent}. For within-subject variability, {\textasciitilde}50{\textpercent} of the mice that performed 2 sessions or more were better explained by multiple strategies across sessions. Our results highlight the importance of not averaging across animals and sessions when analyzing rodent behavior. We are currently analyzing an additional, within-session level of strategy switching, and considering the representation of these various systematic differences in neural activity.}, year = {2021}, note = {50th Annual Meeting of the Society for Neuroscience (Neuroscience 2021)}, slug = {item_3353178}, author = {Findling, C and Dayan, P and Pouget, A} }