与约书亚·本吉奥教授探讨主动学习与 GFlowNets

Active Learning and GFlowNets with Professor Yoshua Bengio

约书亚·本吉奥 Yoshua Bengio · ML Street Talk · 2022-02-22 · 约 93 分钟 · 原视频 ↗

打开互动全文版（中英对照 + 朗读 + 问答）→

本期速览 · Overview

约书亚·本吉奥教授讨论 GFlowNets、主动学习，以及如何在复杂组合空间中高效查询训练数据。

Professor Yoshua Bengio discusses GFlowNets, active learning, and how to efficiently query oracles for training data in complex combinatorial spaces.

要点 · TL;DR

• GFlowNets 替代 MCMC，实现高效采样和概率估计。
  GFlowNets replace MCMC for efficient sampling and probability estimation.
• 采样多样性对探索和避免局部最优至关重要。
  Diversity in sampling is key for exploration and avoiding local optima.
• 抽象和因果结构实现分布外泛化。
  Abstraction and causal structure enable out-of-distribution generalization.

核心观点 · Key points

• GFlowNets 是可学习的 MCMC 替代方案，能高效采样和估计概率。
  GFlowNets are a learnable replacement for MCMC, enabling efficient sampling and probability estimation.
• 采样多样性对于探索和避免强化学习中的局部最优至关重要。
  Diversity in sampling is crucial for exploration and avoiding local optima in RL.
• 抽象和因果结构是分布外泛化的关键。
  Abstraction and causal structure are key to out-of-distribution generalization.
• GFlowNets 可以对组合对象上的复杂分布建模，助力科学发现。
  GFlowNets can model complex distributions over compositional objects, aiding scientific discovery.
• 不确定性建模和主动学习对于高效探索至关重要。
  Uncertainty modeling and active learning are essential for efficient exploration.

反共识 · Contrarian takes

• 与标准强化学习不同，GFlowNets 按奖励比例采样路径，而非最大化奖励。
  GFlowNets sample paths proportional to reward, not maximizing it, unlike standard RL.
• 大型神经网络可能略有意识，但我们缺乏对意识的科学理解。
  Large neural nets may be slightly conscious, but we lack scientific understanding of consciousness.
• 线性与分段线性主导现代机器学习，与平滑非线性方法相反。
  Linearity and piecewise linearity dominate modern ML, contrary to smooth nonlinear methods.
• 神经网络无需人类先验即可发现抽象概念，与普遍看法相反。
  Neural networks can discover abstract concepts without human priors, contrary to common belief.
• 意识可能是注意力和世界模型产生的幻觉，而非独立现象。
  Consciousness may be an illusion from attention and world models, not a separate phenomenon.

本期章节 · Chapters（共 25）

• 引言与事务 Introduction and Housekeeping
• GFlowNets 与主动学习 GFlowNets and Active Learning
• GFlowNets 简介 Introduction to GFlowNets
• 想象机器与高尔顿板类比 Imagination Machine and Galton Board Analogy
• GFlowNets 优势：多样性与鲁棒性 Advantages of GFlowNets: Diversity and Robustness
• GFlowNets 简介 Introduction to GFlowNets
• 高尔顿板类比 Analogy with Galton Board
• 与 AlphaZero 对比及多样性 Comparison with AlphaZero and Diversity
• Bengio 教授背景 Professor Bengio's Background
• GFlowNets 与熵估计 GFlowNets and entropy estimation
• 向 Karl Friston 提问 Question for Karl Friston
• GFlowNets 中的多样性与探索 Diversity and exploration in GFlowNets
• 与多臂老虎机的联系 Connection to multi-armed bandits
• GFlowNets 中的探索与不确定性 Exploration and Uncertainty in GFlowNets
• GFlowNets 与结构学习的潜力 The Promise of GFlowNets and Structure Learning
• 游戏中 GFlowNets 与强化学习对比 GFlowNets vs. Reinforcement Learning in Games
• GFlowNets 用于主动学习与探索 GFlowNets for Active Learning and Exploration
• 因果世界模型与抽象 Causal World Models and Abstraction
• 意识与大型神经网络 Consciousness and Large Neural Networks
• 分段线性近似与抽象 On Piecewise Linear Approximation and Abstraction
• 个人研究历程与 Gary Marcus 共识 Personal Research Journey and Convergence with Gary Marcus
• 致谢与结束语 Appreciation and Closing Remarks
• 关于意识主张的讨论 Discussion on consciousness claims
• 基于图的方法与抽象 Graph-based methods and abstractions
• 神经网络中的抽象 Abstraction in Neural Networks

阅读全文双语转录 →