Schedule of Events | Symposia

Context-Dependent Statistical Learning: Bridging Human Cognition and Neural Network Dynamics

Poster Session A - Saturday, March 29, 2025, 3:00 – 5:00 pm EDT, Back Bay Ballroom/Republic Ballroom

Fleming Peck¹ (fpeck@ucla.edu), Hongjing Lu¹, Jesse Rissman¹; ¹UCLA

We aimed to test whether statistical learning can occur in a context-dependent fashion. Participants performed perceptual judgments on a continuous sequence of abstract visual objects. Unbeknownst to participants, the objects were organized in pairs, such that the first object in a given pair predicted the ensuing object. This temporal associative structure remained constant within a context, but periodically the context would change, and with this change came a new set of temporal associations. Regardless of whether context changes were signaled (via border color) or completely latent (n=50 participants for each condition), participants’ response times for predictable versus unpredictable objects grew faster over time, revealing context-dependent learning. Moreover, a final two-alternative forced-choice (2AFC) test showcased above-chance predictions about which object would come next in a given sequence. To gain insight into the mechanisms supporting this learning, we trained neural network models to predict the next item in the sequence and to complete a final 2AFC task, much like the human participants. Despite no explicit coding of context, we found that gated recurrent units enabled networks to acquire and retain knowledge of both temporal associative structures, in contrast to simple recurrent and feedforward units. Additionally, we found a non-monotonic relationship between weight variance at initialization and task accuracy, with peak performance achieved with modest noise. By applying representational similarity analysis to hidden layer activity patterns, we track the emergence of context sensitivity and relate this to task performance. These results provide clues into how the human brain might achieve context-dependent learning.

Topic Area: EXECUTIVE PROCESSES: Goal maintenance & switching