Unverified50% confidenceFactExact time
JEPA过程无需解码器重建像素，编码器将观测压缩为语义嵌入，预测网络在嵌入空间完成跨时间步状态预测
1
Sources
50%
Confidence
Long-term
Relevance
7/19/2026
First Seen
Sources
LeMario：用超级马里奥训练JEPA世界模型的开源实践
Related Claims
UnverifiedJEPA 借鉴了 BYOL 的动量编码器思想，同时吸收了 MAE 的掩码预测机制，但将重建目标替换为抽象嵌入68% similarUnverifiedJEPA 引入非对称结构，其中目标编码器的梯度通常通过指数移动平均（EMA）更新，而非直接反向传播66% similarUnverifiedJEPA 与自监督学习谱系的关键区别在于明确将'预测'而非'重建'或'对比'作为核心训练信号，且预测目标是经过编码的抽象表示65% similarUnverifiedLeCun在2022年提出的JEPA（Joint Embedding Predictive Architecture）在表征空间中直接预测目标嵌入，而非重构原始输入63% similarUnverifiedThe world model of Ha and Schmidhuber employs a three-module architecture: a visual encoder V responsible for perceptual compression, a recurrent neural network M responsible for temporal memory, and a controller C responsible for decision-making.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/564469API
curl https://kongchang.com/api/v1/knowledge/claims/564469MCP
get_claim(id=564469)