Health & Wellness | 健康知识库

by 大牙哥

如何利用神经科学与AI提升学习能力 | Dr. Terry Sejnowski

How to Improve at Learning Using Neuroscience & AI | Dr. Terry Sejnowski

16分钟阅读

如何利用神经科学与人工智能提升学习能力

摘要

Salk研究所计算神经科学家Dr. Terry Sejnowski阐释了大脑如何通过算法进行学习——尤其是由dopamine驱动的强化学习算法——以及理解这些机制如何能从根本上改善我们获取和保留知识的方式。对话涵盖记忆巩固的神经科学、运动对大脑健康的作用，以及实用工具，包括一门免费的”学会如何学习”在线课程。

核心要点

  • 所有动机都由一条学习规则支配：大脑预测未来奖励，根据预测误差进行更新，并构建”价值函数”——这与AlphaGo所使用的算法完全相同。
  • 认知（外显）学习和程序（练习）学习缺一不可——将练习从教育中剔除，在神经学层面是适得其反的。
  • 非快速眼动睡眠期间的睡眠纺锤波可巩固前一天所学的记忆；纺锤波越多，长期记忆保留效果越好。
  • 运动是促进大脑和认知健康最有力的工具——每日体育活动可补充线粒体功能，并在任何年龄维持认知活力。
  • 负面/惩罚性学习远比奖励性学习更有力——一次极其糟糕的经历就能产生永久性学习（这也是PTSD的基础机制）。
  • 主动、费力的学习优于被动接收——自己解决问题比直接被告知答案更有效，无论对人类还是人工智能皆如此。
  • 已有一门免费的”学会如何学习”课程（由Barbara Oakley联合创建），拥有来自200个国家的400万用户，完全免费，无需数学或词汇基础。
  • 像对待人类一样对待人工智能（礼貌、对话式交流）能产生更好的结果，且认知消耗更少，因为这充分利用了已有的社会脑回路。
  • 认知”速度”至关重要——略微超越本能的阅读或学习节奏，如同间歇训练，可提升记忆保留效果和大脑效率。

详细笔记

大脑的核心学习算法

  • basal ganglia负责学习实现目标所需的动作序列——不仅是运动动作，还包括思维模式。
  • 大脑使用reinforcement learning：预测下一个奖励，与实际发生的情况进行比较，并据此更新synaptic plasticity。
  • 这构建了一个价值函数——终身积累的关于什么对自己有益或有害的知识（例如，知道在餐厅点什么菜）。
  • 这一相同的算法驱动了DeepMind的AlphaGo，使其击败了围棋世界冠军。
  • prefrontal cortex主导该系统的”不行动”一侧——规划、冲动控制和社会行为。它是青春期最后成熟的大脑区域。

两种学习系统

  • 认知（外显）学习：皮层驱动、逐步进行、有意识——如阅读教科书。
  • 程序性学习：皮层下（基底神经节驱动）、自动化、通过重复习得——如学习网球发球或解物理题。
  • 两种系统相辅相成、缺一不可——从教育中移除程序性练习（家庭作业、习题集、训练），会同时削弱两个系统。
  • 仅靠课堂教学是不够的；主动解题才是真正构建技能的地方。

惩罚与奖励在学习中的作用

  • 小而渐进的奖励推动逐步提升。
  • 惩罚的效果要强得多——一次高度负面的经历就能产生永久性的、一次试验即可习得的学习。
  • 这是PTSD的基础：单一创伤事件可终身重塑行为。
  • 社会性惩罚（公开纠正、难堪）是一种强有力的学习信号。

睡眠、记忆巩固与睡眠纺锤波

  • Sleep spindles（持续约1–2秒）在非快速眼动睡眠期间以环形波的形式在皮层间传播。
  • hippocampus回放当日经历，触发睡眠纺锤波，将新记忆”揉入”已有的皮层知识中，而不会覆盖原有内容。
  • UC Irvine的Sara Mednick的研究发现：药物zolpidem（Ambien）可将睡眠纺锤波频率提高一倍，服药后学习的受试者次日早晨能回忆起两倍的内容。
  • 注意：zolpidem会消除服药之后的记忆形成——因此它能巩固过去，但会损害新的记忆编码。
  • 运动可增加睡眠纺锤波密度，尤其在快速眼动睡眠期间，这可能有助于运动学习的巩固。

运动：首要的大脑健康工具

  • Dr. Sejnowski每天在Salk研究所的海滩跑步，并每天攀爬悬崖（340英尺），还在阿尔卑斯山徒步旅行。
  • 运动可补充mitochondrial function——线粒体随年龄增长而衰退，导致细胞能量（ATP）和认知活力下降。
  • High-intensity interval training（HIIT）式间歇训练（例如10秒冲刺）尤其有效——类似于阅读时的认知”间歇训练”。
  • 运动对每个器官系统均有益：心血管、免疫、神经系统。
  • 定期运动被誉为**“你能服用的最好、最廉价的药物”**。

认知储备与阿尔茨海默病预防

  • 一项中国研究发现，Alzheimer’s disease发病最早的是没有接受过正规教育的人，发病最晚的是受教育程度最高的人。
  • 假说：认知锻炼越多 = 神经储备越多——大脑能在症状出现之前承受更多损伤。
  • 这与线粒体储备的概念相似——早期积累越多，可供使用的时间越长。

学会如何学习——免费课程

  • 课程名称：Learning How to Learn
  • 创建者：Dr. Terry Sejnowski 与 Barbara Oakley
  • 形式：MOOC（大规模开放在线课程）——约10分钟的短视频，约50–60节课，历时一个月，包含测验、考试和同伴讨论区。
  • 费用：免费
  • 获取途径：Coursera（原始平台）——搜索”Learning How to Learn”
  • 受众：面向学生设计，但最受欢迎的群体是25–35岁——需要高效学习新技能的在职成年人。
  • 涵盖主题：如何克服拖延、考试焦虑，如何举一反三，主动学习与被动学习策略。
  • 覆盖范围：200个国家400万学员，好评率98%。

人工智能作为学习与认知工具

  • ChatGPT、Claude等大型语言模型（LLM）并非鹦鹉学舌——它们能从训练数据中举一反三，解决新问题，与大脑的泛化方式类似。
  • 关键洞见：以对话方式对待人工智能（礼貌，如同对待人类）能激活已有的社会脑回路，使交互更省认知资源，并产生更好的输出结果。
  • 不同的大型语言模型各有所长——Google Gemini在数学推理方面有显著提升（在链式推理问题上从约20%提升至约80%的准确率）。
  • 人工智能无法替代程序性学习——用AI写歌的学生，并未构建音乐技能所对应的底层神经架构。

大脑连接与发育

  • 婴儿大脑经历突触大量过度生长，随后进行突触修剪——仅保留能量效率最高的连接。
  • 随着年龄增长，皮层变薄，连接性降低，但早期记忆依然牢固，因为它们是在可塑性最强的时期形成的。
  • 新型光学记录技术可同时记录数十个脑区数以万计的神经元活动——揭示出在任何复杂任务中，所有脑区都在相互作用，而不仅仅是”相关”区域。
  • 视觉皮层接收到的输入，来自运动系统的比来自眼睛本身的还要多（UCLA的Mriganka Sur在小鼠中的研究成果）。

涉及概念

  • reinforcement learning
  • basal ganglia
  • dopamine
  • synaptic plasticity
  • value function
  • procedural learning
  • cognitive learning
  • sleep spindles
  • memory consolidation
  • hippocampus
  • prefrontal cortex
  • mitochondrial function
  • PTSD
  • Alzheimer’s disease
  • high-intensity interval training
  • neural plasticity
  • large language models
  • slow wave sleep
  • REM sleep
  • cognitive reserve

English Original 英文原文

How to Improve at Learning Using Neuroscience & AI

Summary

Dr. Terry Sejnowski, computational neuroscientist at the Salk Institute, explains how the brain learns through algorithms — particularly the reinforcement learning algorithm driven by dopamine — and how understanding these mechanisms can dramatically improve how we acquire and retain knowledge. The conversation covers the neuroscience of memory consolidation, the role of exercise in brain health, and practical tools including a free online course on learning how to learn.

Key Takeaways

  • All motivation is governed by a single learning rule: the brain predicts future rewards, updates based on prediction errors, and builds a “value function” — the same algorithm used by AlphaGo.
  • Both cognitive (explicit) and procedural (practice-based) learning are essential — removing practice from education is neurologically counterproductive.
  • Sleep spindles during non-REM sleep consolidate memories learned the previous day; more spindles = better long-term retention.
  • Exercise is the most powerful tool for brain and cognitive health — daily physical activity replenishes mitochondrial function and maintains cognitive vigor at any age.
  • Negative/punishment-based learning is far more powerful than reward-based learning — one very bad experience can produce permanent learning (the basis of PTSD).
  • Active, effortful learning outperforms passive intake — solving problems yourself beats being told the answer, both for humans and AI.
  • A free “Learning How to Learn” course exists (co-created with Barbara Oakley) with 4 million users across 200 countries — zero cost, no math or vocabulary prerequisites.
  • Treating AI like a human (being polite, conversational) yields better results and is less cognitively draining because it leverages existing social brain circuits.
  • Cognitive “velocity” matters — pushing slightly beyond your reflexive reading or learning pace, like interval training, improves retention and brain efficiency.

Detailed Notes

The Brain’s Core Learning Algorithm

  • The basal ganglia is responsible for learning sequences of actions to achieve goals — not just motor actions but also thinking patterns.
  • The brain uses reinforcement learning: it predicts the next reward, compares that to what actually happens, and updates synaptic plasticity accordingly.
  • This builds a value function — a lifetime accumulation of knowledge about what is good or bad for you (e.g., knowing what to order at a restaurant).
  • This same algorithm powered DeepMind’s AlphaGo, which defeated the world Go champion.
  • The prefrontal cortex governs the “no-go” side of this system — planning, impulse control, social behavior. It is the last brain region to mature in adolescence.

Two Types of Learning Systems

  • Cognitive (explicit) learning: cortical, step-by-step, conscious — like reading a textbook.
  • Procedural learning: subcortical (basal ganglia-driven), automatic, acquired through repetition — like learning a tennis serve or solving physics problems.
  • The two systems are complementary and necessary — removing procedural practice (homework, problem sets, drills) from education undermines both systems.
  • Classroom instruction alone is insufficient; active problem-solving is where real skill is built.

Punishment vs. Reward in Learning

  • Small, incremental rewards drive gradual improvement.
  • Punishment is dramatically more effective — one highly negative experience can produce permanent, one-trial learning.
  • This underlies PTSD: a single traumatic event can reorganize behavior for life.
  • Social punishment (public correction, embarrassment) is a potent learning signal.

Sleep, Memory Consolidation & Sleep Spindles

  • Sleep spindles (lasting ~1–2 seconds) travel in circular waves across the cortex during non-REM sleep.
  • The hippocampus replays daily experiences, triggering sleep spindles that “knead” new memories into existing cortical knowledge without overwriting it.
  • Research by Sara Mednick (UC Irvine): the drug zolpidem (Ambien) doubles sleep spindle frequency, and subjects who took it after learning recalled twice as much the next morning.
  • Caveat: zolpidem wipes out memory formation after the drug is taken — so it consolidates the past but impairs new encoding.
  • Exercise increases sleep spindle density, particularly REM sleep, which may support motor learning consolidation.

Exercise as the Primary Brain Health Tool

  • Dr. Sejnowski runs daily on the beach at the Salk Institute and climbs the cliff (340 ft) each day; hikes in the Alps.
  • Exercise replenishes mitochondrial function — mitochondria decline with age, reducing cellular energy (ATP) and cognitive vigor.
  • High-intensity interval training (HIIT)-style intervals (e.g., 10-second sprints) are especially effective — analogous to cognitive “intervals” during reading.
  • Exercise benefits every organ system: cardiovascular, immune, neurological.
  • Regular exercise is described as “the best and cheapest drug you can take.”

Cognitive Reserve and Alzheimer’s Prevention

  • A Chinese study found that onset of Alzheimer’s disease was earliest in those with no formal education and latest in those with the most education.
  • Hypothesis: More cognitive exercise = more neural reserve — the brain can sustain more damage before symptoms appear.
  • This mirrors the concept of mitochondrial reserves — the more you build up early, the longer you can draw on them.

Learning How to Learn — Free Course

  • Course name: Learning How to Learn
  • Created by: Dr. Terry Sejnowski and Barbara Oakley
  • Format: MOOC (Massive Open Online Course) — bite-sized ~10 minute videos, ~50–60 lessons over one month, includes quizzes, tests, and peer forums.
  • Cost: Free
  • Available on: Coursera (originally) — search “Learning How to Learn”
  • Audience: Designed for students but most popular with ages 25–35 — working adults who need to learn new skills efficiently.
  • Topics covered: How to overcome procrastination, exam anxiety, how to generalize knowledge, active vs. passive learning strategies.
  • Reach: 4 million learners across 200 countries, 98% approval rating.

AI as a Learning and Cognitive Tool

  • Large language models (LLMs) like ChatGPT and Claude are not parrots — they generalize from training data to solve new problems, similar to how the brain generalizes.
  • Key insight: Treating AI conversationally (politely, as you would a human) activates existing social brain circuits, making the interaction less cognitively draining and producing better outputs.
  • Different LLMs have different strengths — Google Gemini has improved significantly at mathematical reasoning (from ~20% to ~80% accuracy on chain-of-reasoning problems).
  • AI cannot replace procedural learning — a student who uses AI to write a song is not building the underlying neural architecture of musicianship.

Brain Connectivity and Development

  • Infant brains undergo massive synapse overproduction followed by synaptic pruning — retaining only the most energetically efficient connections.
  • As we age, the cortex thins and connectivity decreases, but early memories remain rock-solid because they were laid down during peak plasticity.
  • New optical recording techniques allow simultaneous recording from tens of thousands of neurons across dozens of areas — revealing that all brain areas interact during any complex task, not just the “relevant” one.
  • More visual cortex input comes from the motor system than from the eye itself (research by Mriganka Sur at UCLA in mice).

Mentioned Concepts

  • reinforcement learning
  • basal ganglia
  • dopamine
  • synaptic plasticity
  • value function
  • procedural learning
  • cognitive learning
  • sleep spindles
  • memory consolidation
  • hippocampus
  • prefrontal cortex
  • mitochondrial function
  • PTSD
  • Alzheimer’s disease
  • high-intensity interval training
  • neural plasticity
  • large language models
  • slow wave sleep
  • REM sleep
  • cognitive reserve

关系图谱