Marvin Maechler¹ (), Ansh Soni¹, Alan Stocker¹; ¹University of Pennsylvania

Decision-making often involves categorical choices that can bias subsequent judgments (e.g., confirmation bias). We have previously shown that when observers categorize a stimulus before estimating it, their estimates are systematically biased. This bias is not due to poor calibration or insufficient motivation because it persists in the presence of detailed feedback (enabling learning) and monetary incentives (enhancing motivation). This suggests that confirmation bias reflects deliberate behavior, although its underlying motivation remains unclear. We used recurrent neural networks to test whether confirmation bias emerges naturally from resource-rational behavior. We artificially imposed memory limitations on the networks by adding varying amounts of noise to every neuron’s activation function at each time step. We trained the networks to perform the same psychophysical task used in our previous experiments on human observers. The task in these experiments was to infer the mean of a stimulus generating distribution based on a sequence of stimulus samples. The observer was asked to categorize the stimuli relative to a decision boundary before estimating the generative mean. Importantly, the networks were trained solely to minimize estimation error without explicit penalty for inconsistency or categorical errors. In the presence of noise, the network model with access to its own past choice learned to rely on it when making an estimate, thereby introducing confirmation bias. Despite the bias, the model achieved lower root mean squared error than the unbiased comparison model without access to its own choice. The performance advantage demonstrates that incorporating the categorical choice as a form of compressed memory preserves some of the information that is otherwise lost due to memory noise. Our results show that confirmation bias can emerge from computational constraints that are ubiquitous to neural systems (i.e., noise). In conclusion, choice-induced bias may reflect efficient neural computation rather than cognitive limitations.

Acknowledgements: This work was supported by the NSF CRCNS grant IIS-1912232 to A.A.S.