20260625:很多新内容
This commit is contained in:
40
concepts/leakage-free-state-prediction.md
Normal file
40
concepts/leakage-free-state-prediction.md
Normal file
@@ -0,0 +1,40 @@
|
||||
---
|
||||
title: "Leakage-Free State Prediction"
|
||||
created: 2026-06-24
|
||||
updated: 2026-06-24
|
||||
type: concept
|
||||
tags: ["jepa", "pretraining", "world-model", "information-leakage"]
|
||||
sources:
|
||||
- "[[vla-jepa-2026]]"
|
||||
---
|
||||
|
||||
# Leakage-Free State Prediction
|
||||
|
||||
Leakage-Free State Prediction 是 VLA-JEPA 的核心设计原则:未来帧仅作为监督目标,永不作为模型输入。这是对当前 latent-action 预训练方法中信息泄漏问题的直接修复。
|
||||
|
||||
## 问题:信息泄漏
|
||||
|
||||
传统 latent-action 方法将当前观察和未来观察同时馈入同一模块,或允许未来上下文影响学习的 latent action。这创造了捷径:latent action 可以简单地编码未来本身,而非捕获状态转移。
|
||||
|
||||
## VLA-JEPA 的方案
|
||||
|
||||
```
|
||||
传统方法(有泄漏): VLA-JEPA(无泄漏):
|
||||
I_t + I_{t+1} → latent action I_t → latent action
|
||||
(未来可影响编码) I_{t+1} → target(stop-grad)
|
||||
latent action → predict target
|
||||
```
|
||||
|
||||
Target encoder 被冻结(stop-gradient),仅用于产生监督信号。Student 无法"看到"未来。
|
||||
|
||||
## 效果
|
||||
|
||||
- Latent action 被迫编码真正的动态信息而非捷径
|
||||
- 对相机运动和背景变化天然鲁棒
|
||||
- 学到的表示更具语义抽象性
|
||||
|
||||
## 参考
|
||||
- [[vla-jepa-2026]]
|
||||
- [[jepa]]
|
||||
- [[information-leakage-vla]]
|
||||
- [[latent-action-pretraining]]
|
||||
Reference in New Issue
Block a user