本文探讨了人类欣赏音乐的神经科学基础，指出音乐品味是一个循序渐进的过程，大脑通过预测旋律和节奏来获得愉悦感。文章以Daniel Levitin的观点为例，阐述了大脑如何运用小脑等区域追踪音乐序列并预测后续发展。当音乐符合或巧妙地打破听众的预期时，大脑的奖励系统会被激活，产生“惊喜”与“解决”的动态平衡，从而带来情感体验。文章还将这一过程与大型语言模型的训练方式进行了类比，并提出了AI在“反应”和“评估”层面可以改进的方向，以实现更具“主动好奇心”和“内在动机”的AI。

🎵 音乐品味是逐步建立的：人类通常从简单易懂的音乐（如流行乐）开始接触，然后随着时间的推移，逐渐深入到更复杂的音乐类型和子流派。这是一个大脑不断学习和构建预测模型的过程。

🧠 大脑的音乐预测机制：根据神经科学家Daniel Levitin的观点，我们的大脑（特别是小脑）会追踪音乐序列并预测接下来的旋律和节奏。音乐带来的愉悦感源于这种“预期”与“惊喜”之间的互动：过于可预测会让人感到无聊，完全不可预测则如同噪音。

✨ 音乐的情感力量：音乐家的艺术性体现在他们如何巧妙地操控听众的预测过程。当音乐在紧张感和释放感之间切换时，例如电子音乐中的“drop”或歌曲的有力副歌，会激活大脑的奖励系统。这种愉悦感来自于预期与实际听到的内容之间的动态变化，包括确认、惊喜和解决。

💡 ITPRA理论解释音乐体验：David Huron的ITPRA理论（想象、紧张、预测、反应、评估）更详细地描绘了这一过程。它包括了我们对音乐走向的想象、预期临近时的紧张感、正确预测的微小奖励、对意外的快速反应以及对音乐结果的意识性反思。

🤖 AI与音乐体验的类比与启示：大型语言模型的训练过程与音乐欣赏的预测与奖励机制有相似之处，都依赖于预测和最小化预测误差。然而，AI在“反应”（如主动提问）和“评估”（如自我反思和寻找洞见）方面仍有提升空间，可以通过引入“主动好奇心”和“内在动机”来改进。

Published on August 15, 2025 5:49 PM GMT

What happens if you play hard techno or Death metal to a person that never listen to music? I would bet they won’t appreciate it. Musical taste is usually built over time. People usually start listening music with something “simple” like radio pop, then they like some songs and dive deeper into a genre, starting with easy pop-like songs, and then discovering lots of sub-genres. Switching from metal to techno is also usually through more predictable slower songs. Our journey through genres is a gradual process of building a predictive model in our brain.

The Musical Brain

Our brains are constantly predicting what comes next in a melody or rhythm. As neuroscientist Daniel Levitin explains in “This Is Your Brain on Music”, our brain uses regions like the cerebellum to track musical sequences and predict what’s next. The pleasure we get from music comes from the interplay between expectation and surprise. A song that is too predictable is boring; one that is completely unpredictable is just noise.

The emotional power of music lies in how an artist skillfully plays with this predictive process. When we experience that thrilling moment of tension and release—like a dramatic “drop” in electronic music or a powerful chorus—our brain’s reward system is activated. The pleasure isn’t just in getting the prediction right; it’s in the dynamic between what we expect and what we get:

Confirmation: When a song resolves as our brain predicts, it provides a sense of comfort and satisfaction.Surprise: When a musician introduces an unexpected chord or a syncopated beat, it violates our expectations, creating a “prediction error” that jolts our attention.Resolution: The true artistry is in the masterful resolution of that surprise. When the unexpected is woven back into a satisfying pattern, our brain experiences a powerful emotional release.

David Huron’s book, “Sweet Anticipation,” offers a more detailed framework for this experience called ITPRA theory:

Imagination: We anticipate where the music might go.Tension: We feel a buildup as an expected moment approaches.Prediction: We get a small reward when our prediction is correct.Reaction: We have a fast, reflexive response to a surprise.Appraisal: We consciously reflect on the musical outcome.

The AI Parallel

This process of prediction, surprise, and reward is exactly how we train Large Language Models. An LLM’s primary goal is to predict the next token based on all previous tokens. Its training is driven by minimizing prediction error using a cross-entropy loss function, which penalizes the model more heavily for being surprised by the correct next token. After this loss is low enough, we usually do another step of reinforcement learning fine tuning, where we ask to generate several complete answers, and select the one that’s closer aligns to what we expect from the model.

LLMs have some of ITPRA theory elements. Inference time compute gives us Imagination, and the context window builds Tension. Prediction is their core function. But they lack a sophisticated Reaction and Appraisal. When faced with a surprising or ambiguous prompt, a model doesn’t get curious; it just makes its most probable guess. It doesn’t appraise its own answers for insight, only for alignment with its training data. Can we do better?

Reaction as Active Curiosity: When an LLM is “surprised” by a confusing prompt, its reaction shouldn’t be to guess. It should be to ask a clarifying question. This turns a passive prediction error into an active search for more information. This is a hot topic, and people build routers to direct different prompts to different models, for exampleAppraisal as Intrinsic Motivation: We can reward a model not just for being “correct,” but for being useful or insightful. By creating a reward signal for reducing uncertainty, we can teach the model to appraise its own lines of reasoning. It would learn to seek the sweet spot between boringly simple and frustratingly complex.

Final thought

If you want someone to like your music, don’t play your favorite song first. You have to let them build a good prediction system. Start with similar music they already like, and gradually introduce songs with more elements from your taste. When your tastes converge, you can shock them with your favorite song

Discuss