Neuroscience experiments have progressively produced larger datasets, demanding analytic approaches that extract dynamics from neural recordings at the population level. While traditional methods focus on the fast (within-trial) components of neural activity, nearly all cognitive processes, including learning, memory, and decision-making, require computations spanning multiple timescales. For example, in a recently designed value-based decision-making task where rats reveal their valuation of a reward based on their willingness to wait for it, rats have been shown to modulate their wait times to uncued changes in reward distributions over blocks of trials. Inactivation of the rat lateral orbitofrontal cortex (lOFC) reduces rats’ sensitivity to these changes, suggesting that this brain area plays a critical role in rats’ sensitivity to long timescale changes in reward statistics. Traditional data-analysis approaches applied to neural recordings from lOFC would be unsuitable for capturing these contextual effects, given their exclusive focus on fast within trial computations. I will present our results using a Hierarchical Kalman Filter (HKF) that explicitly accounts for multiple interacting time scales to fit lOFC Neuropixels recordings, providing estimates of low dimensional latent structure of neural activity within a trial and across trials, separately. This approach is mathematically tractable, allowing us to derive closed-form expressions for inference and parameter estimation (using Expectation Maximization). Our HKF analysis reveals that uncued reward blocks that dictate reward statistics over long timescales are represented in the slow latent dynamics of OFC population responses, and that faster within-trial dynamics reflect single trial reward offers. These findings establish the HKF as a powerful framework for characterizing interpretable, low-dimensional neural dynamics in cognitive processes that involve computations on multiple timescales.
