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

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

From simple to complex: shared learning dynamics in humans and neural networks

Alice Zhang¹, Jirko Rubruck¹, Satwik Bhattamishra¹, Christopher Summerfield¹; ¹University of Oxford

Presenter: Christopher Summerfield

Deep neural networks demonstrate a well-documented simplicity bias—the tendency to learn simple functions before acquiring more complex ones. This bias towards simplicity is thought to enable overparameterized models to successfully generalize to unseen data rather than overfitting to examples seen in training. Complementary work in psychology has demonstrated human simplicity biases in several domains. Here, we aim to unite these two streams by comparing human and neural network simplicity biases side-by-side in a Boolean classification task. We demonstrate that both humans and models initially learn simple rules before mastering a more complex function. We also provide evidence that human learners rely on the simple functions they learned early on to classify out-of-distribution examples, suggesting that dynamical simplicity biases are important for generalization.

Topic Area: Brain Networks & Neural Dynamics

Extended Abstract: Full Text PDF