🔘 Paper page: arxiv.org/abs/2108.08810v1 Abstract «Convolutional neural networks (CNNs) have so far been the de-facto model for visual data. Recent work has shown that (Vision) Transformer models (ViT) can achieve comparable or even superior performance on image classification tasks. This raises a central question: how are Vision Transformers solving these tasks? Are they acting like convolutional networks, or learning entirely different visual representations? Analyzing the internal representation structure of ViTs and CNNs on image classification benchmarks, we find striking differences between the two architectures, such as ViT having more uniform representations across all layers. We explore how these differences arise, finding crucial roles played by self-attention, which enables early aggregation of global information, and ViT residual connections, which strongly propagate features from lower to higher layers. We study the ramifications for spatial localization, demonstrating ViTs successfully preserve input spatial information, with noticeable effects from different classification methods. Finally, we study the effect of (pretraining) dataset scale on intermediate features and transfer learning, and conclude with a discussion on connections to new architectures such as the MLP-Mixer.» Authors Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, Alexey Dosovitskiy Liked this post? Follow this blog to get more. 

The introduction of the cotton gin wasn’t accompanied by an entire genre of Hollywood movies dedicated to the gin “singularity”. Nor did we usher in Golden Age of telecommunications with blockbuster killer “phone web” stories. Artificial Intelligence is different. Like other disruptive technologies, it is having far-ranging effects, good and bad. But uniquely, quality AI information is clouded by the AI apocalypse narrative. If you google the field, you’ll be challenged to separate medical imaging wheat from AGI chaff. (Don’t tell anyone, friends, there’s no magic here, it’s just math.) AI alone is no more likely to take over the world than is your calculator. Well, unless it’s used as a deniability smokescreen: “It’s not my fault the killer robot smashed your house, it was the AI that did it”. Honestly, what worries me most about AGI is the distraction it creates from the real ways that AI can make a massive positive difference in our lives. And the Winter/Summer AI cycle is a massive dampener. AI winter system and status AI hype is nonlinear: a bit of AI hype starts a flywheel effect, often pollinated by well-meaning journalists. AI hype is particularly mutant, which means that those of us...