1 h 1 min
Seminar, Hosted by Center for Brains, Minds and Machines (CBMM), MIT
Aligning deep networks with human vision will require novel neural architectures, data diets and training algorithms
Brown University
Recording
Abstract
Recent advances in artificial intelligence have been driven mainly by the rapid scaling of deep neural networks (DNNs), which now contain unprecedented numbers of learnable parameters and are trained on massive datasets covering large portions of the internet. This scaling has given DNNs visual competencies approaching human levels, yet even the most sophisticated DNNs still show strange, inscrutable failures that diverge markedly from human behavior, a misalignment that seems to worsen as models grow.
This talk presents work addressing this misalignment by developing DNNs that mimic human perception through computational, algorithmic and representational principles fundamental to natural intelligence. It first reviews efforts to characterize human visual strategies in image categorization tasks and to contrast them with modern deep networks, with initial results suggesting that novel data regimens and training algorithms are needed for deep networks to learn more human-like visual representations. It then shows results suggesting that neural architectures inspired by cortex-like recurrent neural circuits offer a compelling alternative to prevailing transformers, particularly for tasks that require visual reasoning beyond simple categorization.