CVPR 2026 ReLearn 工作坊：机器还能向人类学什么
精选理由
Efros、Damen、Manling Li 三个人在 CVPR 2026 聊机器还能从人类学什么，从数据增广聊到空间推理缺口，观点挺尖锐。
CVPR 2026 的 ReLearn 工作坊围绕“机器还能向人类学什么”展开讨论。Alyosha Efros 指出当前 AI 可能已经过度依赖人类设计，连自监督视觉方法都建立在人类设计的增广操作之上。Dima Damen 提倡以自我中心视觉为方向，让模型从单一生命的连续经验中学习。Manling Li 则梳理了基础模型在空间理解上的具体缺口，涵盖从拓扑推理到度量推理的多个层面。
原文 · Lambda
What should machines still learn from humans?
That was the question at CVPR 2026's ReLearn workshop. Alyosha Efros argued today's AI may already learn too much from us, since even self-supervised vision leans on human-designed augmentations. Dima Damen made the case for egocentric vision: learning from the continuous experience of one life. Manling Li traced where foundation models still fail to build spatial understanding, from topological to metric reasoning.
https://t.co/dGfoPJ2ehG
