Yann LeCun explains self-supervised learning as a way for machines to learn background knowledge from observation, like humans do, and why it's crucial for efficient learning.
要点 · TL;DR
自监督学习是智能的暗物质，通过填空学习世界模型。 Self-supervised learning is the dark matter of intelligence, learning world models by filling in blanks.
智能即预测；基于梯度的学习通过可微世界模型实现推理。 Intelligence is prediction; gradient-based learning enables reasoning via differentiable world models.
大多数智能是习得的，而非硬连线；情感是自主系统的组成部分。 Most intelligence is learned, not hardwired; emotions are integral to autonomous systems.
核心观点 · Key points
自监督学习是智能的暗物质，通过观察学习世界模型至关重要。 Self-supervised learning is the dark matter of intelligence, essential for learning world models through observation.
智能本质上是预测；填补空白预测未来、推断过去并构建常识。 Intelligence is essentially prediction; filling in blanks predicts future, infers past, and builds common sense.
基于梯度的学习是关键；通过模拟和规划进行推理使用可微世界模型。 Gradient-based learning is key; reasoning via simulation and planning uses differentiable world models.
大多数智能是习得的，而非硬连线；内在驱动力是天生的，但世界模型是学来的。 Most of intelligence is learned, not hardwired; intrinsic drives are innate, but world models are learned.
自主智能系统将把情感作为其架构的组成部分。 Autonomous intelligent systems will have emotions as integral part of their architecture.
反共识 · Contrarian takes
仅靠文本不足以实现 AGI；视觉提供了关于世界的更多信息。 Text alone is insufficient for AGI; vision provides much more information about the world.
意识可能源于只有一个世界模型，这是一种局限，而非特性。 Consciousness may arise from having only one world model, a limitation, not a feature.
像 Barlow Twins 这样的非对比方法比对比学习更有前景用于自监督视觉。 Non-contrastive methods like Barlow Twins are more promising than contrastive learning for self-supervised vision.
社交媒体的负面影响被夸大；研究表明与两极分化没有因果关系。 Social media's negative effects are overstated; studies show no causal link to polarization.
中文房间论证是荒谬的；通过查找表模仿智能并非真正的智能。 The Chinese room argument is ridiculous; mimicking intelligence through a lookup table is not intelligence.
本期章节 · Chapters（共 46）
引言与自监督学习定义Introduction and Self-Supervised Learning Definition
背景知识与学习效率Background Knowledge and Learning Efficiency
自监督学习信号对比其他范式Signal in Self-Supervised Learning vs Other Paradigms
填补空白：语言与视觉Filling the Gap: Language and Vision
自监督学习即填空Self-Supervised Learning as Filling in the Blanks
视觉与语言的差异Differences Between Vision and Language
语言模型的独立性假设Independence Assumption in Language Models
批评：仅是统计，非智能Criticism: Just Statistics, Not Intelligence
类比：大脑高低级层次Analogy: High-Level vs Low-Level in Brain
机器学习挑战：世界模型、推理与分层规划Challenges in Machine Learning: World Models, Reasoning, and Hierarchical Plans
驾驶如同异步象棋Driving as an Asynchronous Chess Game
智能暗物质与猫级知识Dark Matter of Intelligence and Cat-Level Knowledge
自监督学习与迁移学习Self-supervised learning and transfer learning
数据增强与对比学习Data augmentation and contrastive learning
非对比方法与互信息Non-contrastive methods and mutual information
物理理解与先天后天Physics understanding and nature vs nurture
数据增强与掩码Data augmentation and masking
从视频中学习与数据增强Data Augmentation and Learning from Video
从手工到端到端学习Evolution from handcrafted to end-to-end learning
意识是大脑局限的产物Consciousness as a consequence of brain limitations
争议观点：学习 vs 硬编码Ideas that many disagree with: learning vs. hardwiring
人类对死亡的认知Human awareness of mortality
对死亡与动机的反思Reflections on Death and Motivation
论意识与智能实体On Consciousness and Intelligent Entities
机器人民权与死亡Civil Rights for Robots and Mortality
个性化 AI 的伦理法律问题Ethical and legal issues with personalized AI
中文屋论证与强 AIChinese room argument and strong AI
创造紧迫感与激励他人Creating Urgency and Inspiring People
FAIR 八周年：成功与教训FAIR's 8th Anniversary: Successes and Lessons
元宇宙与 AR/VRThe Metaverse and AR/VR
为 Facebook 辩护Defending Facebook Against Criticism
平衡正负影响Balancing Positive and Negative Effects
Meta 的技术挑战Technological Challenges at Meta
科技领袖 AI 焦点讨论Discussion on AI Focus of Tech Leaders
ICLR 拒稿与审稿过程Paper Rejected from ICLR and Review Process
非对比联合嵌入方法Non-contrastive joint embedding methods
VICReg 论文背后的故事Story behind the VICReg paper
同行评审与会议Peer review and conferences
论文评审系统与声誉Paper Review System and Reputation
涌现与复杂系统Emergence and Complex Systems
复杂性与涌现Complexity and Emergence
构建表现力电子管乐器Building an Expressive Electronic Wind Instrument