Network dimensions alter reversal learning strategies

How do animals adapt in a dynamic world? Sometimes, our expectations about the world are drastically violated, such as when we accidentally bite into spoiled food. In these circumstances, we must adjust or face further harm. A classical experimental paradigm to study this behavior is reversal learning, where the values of stimuli are suddenly reversed, for example from positive to negative. Here we investigate how computational models of neural circuits can express the many behavioral strategies empirically observed in reversal learning. We report that simply introducing more neurons in a neural network can dramatically alter an agent's reversal learning strategy. We first demonstrate this phenomenon in simple few-parameter actor-critic models trained on a reversal learning task. Introducing one additional parameter into a two-parameter model can induce a switch in its learning strategy, even though the original strategy remains available to the model. Next, we show that this phenomenon occurs more generally in neural networks trained on the same task. Increasing the width of a neural network (number of neurons per layer) can once again induce a switch in its learning strategy. Moreover, we show that a shallow network and a deep network of equal size can employ different strategies. Together, these results suggest that network dimensions (width, depth) provide an inductive bias towards different learning strategies. Our findings suggest a new perspective to consider in the neural basis of reversal learning.