⌘K

Change language Switch ThemeSign In

Narrow Mode

The Feynman Technique Is Artificial Backpropagation

The Feynman Technique Is Artificial Backpropagation

@vista8

费曼学习法就是人工的反向传播。

读懂一个东西不够，还得用自己的话讲出来。

讲不清楚，说明梯度没有回传到底层，只是在表面记住了token的排列顺序，权重没更新。

费曼方法在测试梯度到底传到了哪一层。

讲不出来，说明只动了表层。

讲得出来但不够简单，说明传到了中间层。

能让一个小孩听懂，说明底层权重已经彻底更新了。

死记硬背是只训练最后一层，融会贯通是全网络微调。Show More

Mar 20, 2026, 11:25 PM View on X

5 Replies

16 Retweets

112 Likes

7,215 Views  向阳乔木 @vista8

One Sentence Summary

The Feynman Technique is analogized to artificial backpropagation in neural networks, explaining learning depth through weight updates.

Summary

The author presents a clever analogy: the Feynman Technique is essentially artificial backpropagation. Merely understanding something isn't enough—you need to explain it in your own words. If you can't explain it clearly, it means the gradient hasn't propagated back to the foundational layer; you've only superficially memorized the token sequence. The Feynman method tests where the gradient has reached: if you can't explain it, only the surface layer was activated; if you can explain it but not simply enough, it reached the middle layer; if you can make a child understand, the foundational weights have been completely updated. Rote memorization trains only the last layer, while deep understanding is full-network fine-tuning. This is a highly insightful analogy that perfectly links core deep learning mechanisms with classic learning methods.

AI Score

89

Influence Score 31

Published At Yesterday

Language

Chinese

Tags

Feynman Technique

Backpropagation

Deep Learning

Learning Methods

Weight Update HomeArticlesPodcastsVideosTweets

The Feynman Technique Is Artificial Backpropagation | Bes...