• Used neuroscience-inspired techniques to understand and improve deep neural networks.• Investigated properties of learned network solutions which are predictive of generalization ability, both in individual networks and across groups of networks.• Developed new behavioral and introspective methods to measure, understand, and improve abstraction abilities in deep neural networks.

• Developed an evidence accumulation task for head-restrained mice in virtual reality and used calcium imaging of activity in large neuronal populations along with novel computational analyses to study the neuronal population activity dynamics underlying decision-making.• Using machine learning approaches that had not been applied to neuroscience data sets previously, we identified multiple major features of the population activity, including a distributed code among heterogeneous individual neurons, history signals for past events lasting seconds, and structured trial-trial variability that was predictive of past and future activity patterns.• Together our results provide data that contradict the long-standing models of evidence accumulation and that propose a novel model of computation based on large-scale neuronal population dynamics.