文章深入探讨了深度学习训练与人类学习之间的类比，并指出其不准确之处。文章认为，将深度学习训练比作人类学习是不恰当的，因为人类学习在很大程度上依赖于先天能力和极少的数据，而在某些方面，如语言和数学的学习上，人类的泛化能力远超深度学习模型。此外，人类大脑在非刻意思考层面并非通用的学习机器，许多认知能力似乎是天生的。文章进一步指出，人类心智的构建主要归功于进化，而非人类的学习过程本身。同样，深度学习模型的构建主要归功于深度学习过程本身，而非训练代码的编写。文章最后探讨了深度学习训练与生物进化的类比，并强调了两者在优化过程上的相似性。

🧠 人类学习与深度学习训练存在显著差异：人类在少量数据下高效学习，而深度学习模型则依赖海量数据。人类的泛化能力在某些领域超越深度学习模型。

🧬 人类认知能力多样性：许多认知能力如绝对音感和心理视觉似乎是先天存在的，并且可能无法通过后天学习获得，这显示了人类大脑的复杂性。

🤔 心智构建的功劳归属：人类心智的构建主要归功于进化，深度学习模型的心智构建主要归功于深度学习过程本身，而非训练代码的编写。

🎯 深度学习训练与生物进化的类比：尽管存在差异，但两者在优化过程中存在相似性，这有助于理解内部不一致性的存在。

Published on February 2, 2025 8:59 PM GMT

A more correct but less concise statement of the analogy might be DL training : DL training code :: human learning : human genome, read as "DL training is to DL training code what human learning is to the human genome". This is sometimes contrasted with an alternative analogy DL training : DL-based AGI :: evolution : human mind.

Why the analogy is bad

Human learning mostly doesn't look much like DL training:

Above the level of basic sensory stuff, humans have very little data to work with
And are correspondingly vastly more sample efficient at learning some sorts of things, such as theory of mind, language, even mathContrast this to LLMs, which learn language and math through terabytes upon terabytes of examples (but currently do not manage to generalize as far as humans in their applications of these things)
Human brains don't seem like terribly general learning machines below the level of deliberate thought (much less general than e.g. transformers)
A lot of cognitive abilities that vary between humans seem innate; they're readily picked up by those with the potential and remain out of reach to those without
E.g. perfect pitch, mental visualization (see aphantasia), etc.
To say nothing of the things no human can learn to do on a basic levelBrain organoids haven't managed to do anything more impressive than pong as far as I know

The higher level problem with this analogy is how it misleadingly/incorrectly assigns credit to processes for the work of building minds. To me, it's clear from the above points that human learning is doing very little of the work of building human minds; evolution did most of that work. On the other hand, the writing of DL training code (at least in the current paradigm) does very little of the work in building the mind of a hypothetical DL-based AGI: the DL process itself is doing most of the work.

What about the "DL training == evolution" analogy?

DL training also doesn't look much like biological evolution in some ways (e.g. much less of an information bottleneck). The reason this analogy works for the specific purpose of establishing an existence proof of inner misalignment with an outer optimizer is that it correctly identifies the optimization process which is doing the bulk of the actual mind-building work in each case.

Discuss