Research Papers 论文研究 5h ago Updated 1h ago 更新于 1小时前 44

Fast Weight Attention for Continual Learning 快速权重注意力用于持续学习

Recurrent fast-weight memories and selective state-space models compress expanding contexts into fixed-size recurrent states, treating state transitions as online learning rules under read-after-write autoregressive semantics The prefix-aligned pair $(\mathbf{x}_t,\mathbf{y}_t)=(\phi(\mathbf{k}_{t-1}),\mathbf{v}_t)$ is identified as the local fast-memory example at step $t$, while same-step associations optimize a different internal objective Normalized first-order updates are derived for both s 循环快速权重记忆和选择性状态空间模型将不断扩展的上下文压缩为固定大小的循环状态,在读写自回归语义下将状态转移视为在线学习规则 前缀对齐对 $(\mathbf{x}_t,\mathbf{y}_t)=(\phi(\mathbf{k}_{t-1}),\mathbf{v}_t)$ 被识别为第 $t$ 步的局部快速记忆样本,而同一步关联则优化不同的内部目标 针对平方误差回归和负内积目标分别推导了归一化的一阶更新规则,产生了 Falcon 系列:Falcon-1(标量 NLMS)、Falcon-2(逐列扩展)和 Falcon-3(滑动窗口小批量),以及相应的内积变体(Falcon-1A/2A/3A) 该框

58
Hot 热度
72
Quality 质量
62
Impact 影响力

Analysis 深度分析

TL;DR

  • Recurrent fast-weight memories and selective state-space models compress expanding contexts into fixed-size recurrent states, treating state transitions as online learning rules under read-after-write autoregressive semantics
  • The prefix-aligned pair $(\mathbf{x}_t,\mathbf{y}t)=(\phi(\mathbf{k}{t-1}),\mathbf{v}_t)$ is identified as the local fast-memory example at step $t$, while same-step associations optimize a different internal objective
  • Normalized first-order updates are derived for both squared-error regression and negative inner-product objectives, yielding the Falcon family: Falcon-1 (scalar NLMS), Falcon-2 (per-column extension), and Falcon-3 (sliding-window mini-batch), with corresponding inner-product variants (Falcon-1A/2A/3A)
  • The framework supports recurrent, masked-parallel, and chunk-parallel implementations with numerically stable positive-decay renormalization
  • Representative variants remain competitive in language modeling and improve length extrapolation on variable-digit addition tasks

Why It Matters

This work provides a principled theoretical framework for understanding how recurrent fast-weight mechanisms enable continual learning in sequence models, bridging the gap between online learning theory and practical recurrent architectures. For AI practitioners building systems that need to adapt over time without catastrophic forgetting, the Falcon family offers concrete, implementable update rules with proven numerical stability. The separation of temporal alignment, plasticity, forgetting, and bounded rehearsal gives researchers a new lens for analyzing and designing adaptive recurrent models.

Technical Details

  • Fast-weight memory formulation: The paper studies state transitions as online learning rules where the recurrent state encodes an expanding context. Under read-after-write autoregressive semantics, the local fast-memory example at step $t$ is the prefix-aligned pair $(\phi(\mathbf{k}_{t-1}), \mathbf{v}_t)$, distinct from the causal same-step association $(\phi(\mathbf{k}_t), \mathbf{v}_t)$ which optimizes a different objective.
  • Falcon update family: Three regression variants are derived—Falcon-1 uses a scalar Normalized Least Mean Squares (NLMS) update, Falcon-2 extends this per-column, and Falcon-3 employs a sliding-window mini-batch approach. Corresponding inner-product variants (Falcon-1A/2A/3A) are derived for negative inner-product objectives.
  • Implementation forms: The framework provides recurrent, masked-parallel, and chunk-parallel computational forms, along with numerically stable positive-decay renormalization to prevent state explosion.
  • Benchmarks: Evaluated on language modeling tasks and variable-digit addition for length extrapolation, with representative variants showing competitive performance and improved extrapolation capabilities.
  • Theoretical decomposition: The framework explicitly separates four key concerns in recurrent sequence models: temporal alignment, plasticity, forgetting, and bounded rehearsal.

Industry Insight

  • The Falcon family provides ready-to-implement online learning rules that could be integrated into production recurrent models requiring continual adaptation, particularly in scenarios where retraining from scratch is prohibitively expensive.
  • The theoretical separation of plasticity and forgetting mechanisms offers a roadmap for addressing catastrophic forgetting in large-scale language models without sacrificing generative performance, a critical challenge for deploying persistent AI agents.
  • The improved length extrapolation on variable-digit addition suggests fast-weight attention mechanisms could enable more robust handling of out-of-distribution sequence lengths, valuable for real-world applications where input lengths vary significantly from training data.

摘要

循环快速权重记忆和选择性状态空间模型将不断扩展的上下文压缩为固定大小的循环状态,在读写自回归语义下将状态转移视为在线学习规则
前缀对齐对 $(\mathbf{x}_t,\mathbf{y}t)=(\phi(\mathbf{k}{t-1}),\mathbf{v}_t)$ 被识别为第 $t$ 步的局部快速记忆样本,而同一步关联则优化不同的内部目标
针对平方误差回归和负内积目标分别推导了归一化的一阶更新规则,产生了 Falcon 系列:Falcon-1(标量 NLMS)、Falcon-2(逐列扩展)和 Falcon-3(滑动窗口小批量),以及相应的内积变体(Falcon-1A/2A/3A)
该框架支持循环、掩码并行和块并行实现,并具备数值稳定的正衰减重归一化
代表性变体在语言建模中保持竞争力,并在变位数加法任务上改善了长度外推能力

深度分析

一句话总结

  • 循环快速权重记忆和选择性状态空间模型将不断扩展的上下文压缩为固定大小的循环状态,在读写自回归语义下将状态转移视为在线学习规则
  • 前缀对齐对 $(\mathbf{x}_t,\mathbf{y}t)=(\phi(\mathbf{k}{t-1}),\mathbf{v}_t)$ 被识别为第 $t$ 步的局部快速记忆样本,而同一步关联则优化不同的内部目标
  • 针对平方误差回归和负内积目标分别推导了归一化的一阶更新规则,产生了 Falcon 系列:Falcon-1(标量 NLMS)、Falcon-2(逐列扩展)和 Falcon-3(滑动窗口小批量),以及相应的内积变体(Falcon-1A/2A/3A)
  • 该框架支持循环、掩码并行和块并行实现,并具备数值稳定的正衰减重归一化
  • 代表性变体在语言建模中保持竞争力,并在变位数加法任务上改善了长度外推能力

为何重要

这项工作为理解循环快速权重机制如何使序列模型实现持续学习提供了严谨的理论框架,弥合了在线学习理论与实际循环架构之间的鸿沟。对于构建需要随时间适应而无需灾难性遗忘的系统的 AI 从业者而言,Falcon 系列提供了具体可实现的

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Research 科学研究 Training 训练 LLM 大模型 Inference 推理