This project aims to build AI systems that match the efficiency of biological vision while preserving state-of-the-art performance. Inspired by how the brain conserves energy, the team will design new neural architectures that use computation selectively and stop early when confident—dramatically reducing unnecessary processing. These models will learn when and where to allocate computation through conditional execution and iterative refinement, improving only the parts of an input that need it.

Beyond engineering gains, the project will test whether these efficient models better reflect neural processing in humans and primates, including whether harder images naturally require more computation—mirroring human reaction times. Ultimately, the team seeks AI models that are both far more efficient and more biologically grounded, with performance validated across visual and generative tasks.