François Chollet 用一个四帖线程解释了他所说的“符号学习”：它仍然是从样本中自动学习映射的机器学习，但学习到的表示是离散、显式、简约且接近代码的，而不是连续曲线式的参数表示。

第一帖：定义表示基底

“Symbolic learning” is simply “machine learning,” but where the substrate is symbolic: the functions you learn are code-like, not curve-like.

“符号学习”就是机器学习，只是它采用符号化基底：学习到的函数更像代码，而不是曲线。

作者补充，符号学习并不排斥数字或概率；关键是学习表示具有离散、显式、简约和代码式的结构。他把这一概念与“参数化学习”或“曲线拟合”相对照。

第二帖：不是手写符号系统

Symbolic learning is emphatically a form of machine learning: the representations are automatically derived from data, just like with deep learning.

符号学习明确属于机器学习：与深度学习一样，它的表示也是从数据中自动获得的。

这里的“符号学习”不等同于 1970—1990 年代由人手工编写表示的“符号 AI”。如果开发者亲自依据样本编写映射函数，作者认为那应称为编程。

第三帖：与深度学习的关系

The goals are the same, the general principles are the same. The representations differ. But they may end up converging eventually.

两者目标相同、一般原则相同；表示形式不同，但最终也可能趋于融合。

作者认为，正确实现的符号学习可与深度学习逐项对应；两者主要差别在表示基底及其相应的学习机制。

第四帖：编码 Agent 与程序合成

In fact what you’re doing is specifically a form of DL-guided program synthesis (that’s what LLM codegen is).

实际上，这具体属于深度学习引导的程序合成——也就是大语言模型代码生成所做的事情。

作者把“让编码 Agent 根据若干输入输出样本编写映射程序”视为一种成本很高的符号学习，并表示还存在数据和计算效率高出多个数量级的方法；该线程没有进一步列出这些方法或量化比较。

作者：François Chollet（@fchollet）
日期：2026 年 9 月 10 日
原始 X：François Chollet on X: ""Symbolic learning" is simply "machine learning" (automatically learn function x -> y given examples of (x, y) pairs), but where the substrate is symbolic, i.e. the functions you learn are code-like, not curve-like. The term is in opposition to "parametric learning" or "cur… / X
线程续帖：François Chollet on X: "Symbolic learning also does not mean "writing the representations by hand" (you are thinking of "symbolic AI" from the 1970s-1990s). Nowadays when you write by hand a function x -> y given examples of (x, y) pairs, we just call that "programming." Symbolic learning is empha… / X
线程续帖：François Chollet on X: "Symbolic learning, correctly implemented, is in every way parallel to deep learning and comparable to it 1:1, but with a different representation substrate, and accordingly with a different learning mechanism. The goals are the same, the general principles are the same. The… / X
线程续帖：François Chollet on X: "If you ask a coding agent to write a program that maps x to y given some (x, y) examples, you're doing a very expensive form of symbolic learning. In fact what you're doing is specifically a form of DL-guided program synthesis (that's what LLM codegen is). However there are… / X