Abstract

Learning general visual representations, those useful for many tasks, is a key challenge in Computer Vision. For many years, Convolution Neural Networks (CNNs), typically trained on the ImageNet dataset, have been used as a “backbone”, or starting point, for downstream tasks. Perhaps surprisingly, recent work has demonstrated large CNNs transfer well to small downstream tasks, even in the few-shot regime. However, such CNNs need large datasets for pre-training. While CNNs appear to have just the right inductive biases for small or medium-scale training, are they still optimal in modern transfer learning regimes? In this talk we will discuss some recent trends, and surprising findings, in architecture design, scaling laws, and visual representation learning.Learning general visual representations, those useful for many tasks, is a key challenge in Computer Vision. For many years, Convolution Neural Networks (CNNs), typically trained on the ImageNet dataset, have been used as a “backbone”, or starting point, for downstream tasks. Perhaps surprisingly, recent work has demonstrated large CNNs transfer well to small downstream tasks, even in the few-shot regime. However, such CNNs need large datasets for pre-training. While CNNs appear to have just the right inductive biases for small or medium-scale training, are they still optimal in modern transfer learning regimes? In this talk we will discuss some recent trends, and surprising findings, in architecture design, scaling laws, and visual representation learning.