Files
myWiki/concepts/intrinsic-dimension.md

37 lines
1.4 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

---
title: "Intrinsic Dimension: 参数空间的内在维度"
created: 2026-06-25
updated: 2026-06-25
type: concept
tags: [dimensionality-reduction, loss-landscape, manifold-learning]
sources: ["[[sen-mapping-networks]]"]
---
# Intrinsic Dimension (内在维度)
Intrinsic Dimension (ID) of objective landscapes 指深度网络损失函数实际依赖的有效参数维度——通常远小于名义参数数量 P。
## 关键研究
Li et al. (2018) 发现:通过在随机低维子空间中优化(而非完整 P 维空间),深度网络仍能达到接近全参数训练的性能。这意味着:
$$\text{ID} \ll P$$
例如,某些任务上仅需几百个参数方向即可取得 90%+ 精度。
## 与 Weight-Manifold Hypothesis 的关联
ID 研究为 [[weight-manifold-hypothesis|Weight-Manifold Hypothesis]] 提供了间接经验证据:
- 若参数空间的有效维度远小于 P则参数很可能位于低维子流形上
- 这意味着训练中实际"探索"的自由度远小于参数总数
## Mapping Networks 如何利用 ID
Mapping Networks 的**隐向量维度 d** 本质上是架构对 ID 的显式估计。通过在 d 维空间优化而非 P 维,直接将内在维度的理论洞察转化为架构设计。
## 参考
- Li et al., "Measuring the Intrinsic Dimension of Objective Landscapes", ICLR 2018
- [[loss-landscape]]
- [[manifold-hypothesis]]