Search Papers | Poster Sessions | All Posters
Poster C80 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
Reinforcement Learning Over Complex Naturalistic Scenes
Daniel B. Ehrlich1 (), Anne G.E. Collins1; 1University of California Berkeley
Significant strides have been made in understanding reward-based learning in neuroscience and psychology. However, many experimental paradigms employ overly simplistic stimuli, diverging from real-life experiences. Here, we investigate how humans learn over well- controlled naturalistic stimuli from the ImageNet dataset. We used a pre-trained neural network to partition scenes into categories of visual and semantic features, drawn from differing levels of abstraction. Recent research highlights the intricate interaction between learning and cognitive processes, like working memory, attention, and perception. We hypothesized that the level of feature abstraction might impact recruitment of cognitive processes during learning. Indeed, we found that far from learning parameters being constant across these varied types of scene constructions, the types of features used impacted overall performance and the learning parameters drawn from best fit models.
Keywords: Reinforcement Learning Working Memory Artificial Neural Networks Cognition