Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session C: Friday, August 15, 2:00 – 5:00 pm, de Brug & E‑Hall

Simultaneous modeling of behavior and dopamine with disentangled RNNs

Siddhant Jain¹, Nathaniel D. Daw², Kim Stachenfeld³, Kevin J Miller⁴; ¹Google, ²Princeton University, ³Columbia University, ⁴University College London, University of London

Presenter: Siddhant Jain

Understanding how neural activity relates to cognitive processes during learning requires methods that can jointly model brain signals and behavior. Here, we extend Disentangled RNNs (DisRNN), an approach for using constrained recurrent neural networks to discover cognitive models, to the case of jointly modeling behavioral data and measurements of neural activity. We augment a DisRNN trained on choice prediction with a separate subnetwork to predict neural activity. We apply this approach to datasets from a simple reward-learning task, consisting of choices, rewards, and a scalar measure of dopamine responses to reward. First, using synthetic data from a Q-learning agent, we demonstrate the approach is able to capture both choices and reward prediction errors with a single set of internal variables, consistent with the groundtruth. Next, we apply this approach to laboratory data from mice performing a similar task, successfully modeling both choice behavior and nucleus accumbens dopamine responses. Analysis of the fit DisRNNs confirms that the same interpretable latent variables are utilized for both choice prediction and dopamine signal prediction, demonstrating the model's potential to uncover cognitive models that bridge behavior and neural data through shared representations.

Topic Area: Reward, Value & Social Decision Making

Extended Abstract: Full Text PDF