NeuroAI

We use artificial neural networks as explicit, testable hypotheses about biological vision, and we use their failures to identify principles that may improve both neuroscience and AI. Our multi-scale model of mouse primary visual cortex integrates structural and physiological measurements, our work aligning in vitro, in vivo, and in silico cell classes tests whether models preserve meaningful biological organization, and our benchmark of unsupervised learning algorithms against human learning evaluates models beyond static task accuracy. We now ask which objectives and architectures produce brain-like selectivity and invariance, whether predictions generalize across animals, stimuli, and states, and which biological principles improve continual learning, efficiency, and interpretability in AI systems.