Brain-like variability in convolutional neural networks reveals evidence-, uncertainty- and bias-driven decision-making

Even when stimuli and tasks are held constant, brain activity fluctuates markedly across trials, yet it is not well understood how these fluctuations affect decision-making and behavior. Here we address this gap in knowledge using a convolutional neural network (CNN) trained on a perceptual decision-making task. Applying an established analytical framework, we show that CNN activity exhibits trial-level variability reflecting findings from human neuroimaging.Examination of the network’s internal state revealed three distinct activation patterns reflecting: (i) decisions dominated by strong sensory evidence, (ii) decisions under low discriminability with weak or ambiguous evidence, and (iii) decisions in which sensory evidence was opposed by bias. Notably, choice bias shifted the decision boundary, improving performance despite low or conflicting evidence. Together, these findings show that variability in CNN activity can serve as a model for understanding how the brain transforms changes in internal states into adaptive behavior, providing a bridge between fluctuations in activity, decision-making, and behavior.