Sensory encoding and shared state

Neural responses vary from trial to trial even when the same stimulus is presented under the same task conditions. However, this variability is not simply noise: some of it can be explained by measured variables such as movement and task structure, and some of it appears to reflect population-wide state that is only visible at the level of many neurons at once. Work on attention and population gain has shown that these shared fluctuations can be structured rather than random (Rabinowitz et al., 2015), and latent variable models have shown that they can often be captured with a small number of modulators (Whiteway et al., 2019). It is therefore still unclear how much of the structure in V1 activity can be explained by observed variables, such as locomotion and task-independent movement, and how much must be inferred from the neural population itself (Musall et al., 2019)(Stringer et al., 2019).

To address this, I use two complementary modeling approaches. First, I fit generalized linear models, which test how much of each neuron’s activity can be explained by measured visual, task, and behavioral variables. This approach is motivated by work showing that movement variables can explain a large fraction of trial-to-trial neural variability during behavior (Musall et al., 2019), and by methods that explicitly separate stimulus-, movement-, and decision-related population signals (Kobak et al., 2016). I expect it to reveal that observed variables such as locomotion or task-independent movement account for part of how sensory information is represented in V1 while still leaving important fluctuations unexplained (Yin et al., 2025). Second, I fit latent variable models, which estimate hidden signals shared across many neurons and can capture population structure that is not obvious from measured regressors alone (Whiteway et al., 2019). My hypothesis is that this latent shared state will help explain residual variability in sensory encoding, particularly during movement states associated with weaker performance (Rabinowitz et al., 2015)(Dadarlat & Stryker, 2017). Ultimately, this project asks whether observed variables and latent population state provide complementary explanations for how movement shapes visual coding.