Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster C in Poster Session C: Wednesday, August 5, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Emergent Superstitious Learning Behaviors and Computations in Neuromodulated Plastic Neural Networks

Yuhao Jin¹, Kenneth Kay², Sam Feng³, Massimo Silvetti⁴, Ferrera, Vincent P.¹, Jacqueline Gottlieb¹; ¹Columbia University, ²Howard Hughes Medical Institute, ³Paris-Sorbonne University Abu Dhabi, ⁴National Research Council

Presenter: Yuhao Jin

Animals continually learn about predictable relationships. However, such “learnable” relationships are often intermixed with observations that are in fact random, which are thus “unlearnable.” It is not known whether predictable and random relations can be differentiated when presented together, such that more cognitive effort can be invested in the former. Previous work presented subjects with “learnable” sets in which the stimuli were implicitly ordered and “unlearnable” sets where stimuli were unordered and feedback was random. Results showed that subjects learned the correct order in the learnable sets but also imposed “superstitious” ordering of the unlearnable sets. Here, to gain more insight on putative neural mechanisms, we meta-trained a plastic recurrent neural network (p-RNNs) that uses neuromodulated Hebbian plasticity to learn autonomously across trials. Intriguingly, p-RNNs also exhibited subjective ordering when tested with unlearnable sets. Further, we found that p-RNNs expressed a reward-sensitive neuromodulatory signal that predicted superstitious ordering and, further, that real and superstitious orderings were represented differently in memory-related population activity. Overall, this work provides an initial point of connection between superstitious learning and lower-level neurocomputational mechanisms.

Topic Area: Memory, Learning & Knowledge Structures