活线脑：神经可塑性、适应性与智能的未来
摘要
神经科学家 David Eagleman 在讨论其著作 Livewired 时提出，大脑与传统的硬件/软件系统有着本质区别——它会在整个生命过程中根据经历不断进行物理层面的重组。对话涵盖了neuroplasticity的原理、brain-computer interfaces的局限性、智能的本质，以及生物大脑能为构建更优 AI 系统提供哪些启示。
核心要点
- 大脑是”活线”的，而非仅仅是”可塑”的 —— 它从不停止重新布线；与被塑形后永久保持形状的塑料物品不同，大脑没有最终固定的状态。
- 不同脑区的可塑性窗口期各异 —— 视觉皮层固化相对较快；运动皮层和躯体感觉皮层则保持较高的可塑性，因为身体处于持续变化之中。
- 可塑性随年龄下降，部分原因在于动机不足，而非单纯的生理因素 —— 保持认知活跃和社交活跃的老年人能够维持显著的neuroplasticity，即便存在阿尔茨海默症的器质性病变。
- 新奇感与挑战是驱动大脑改变的主要因素 —— 日常惯例会降低可塑性；被迫打破习惯性模式（如 2020 年的经历）则会积极促进神经重连。
- 大脑是一台相关性过滤机器 —— 它编码的是对生存和个人目标重要的信息，而非平等对待所有输入的数据。
- 大脑能”说”任何感觉方言 —— 人工耳蜗和视网膜植入之所以有效，并非因为它们完美复制了生物信号，而是因为大脑能自行解读任何传入的数据。
- 内群体/外群体偏见是一种低层级的神经学反应，而非纯粹的认知或文化反应——在标签被分配后仅数分钟内，任意性的标签就能触发这种反应。
- “即时”学习优于死记硬背 —— 由好奇心驱动的情境化学习，比抽象的、脱离语境的教育能产生更强的神经编码。
- AI 若要接近人类水平的智能，很可能需要生存驱动力和相关性感知，而不仅仅是更多的参数或更大的训练集。
详细笔记
什么是”活线”（Livewired）？
Eagleman 创造了”活线”（livewired）这一术语，以取代传统的neuroplasticity框架。“可塑性”（plastic）一词（由 William James 在 100 多年前提出）暗示一个系统被塑形后便保持该形状——这对于塑料制造业而言是准确的，但对大脑而言却具有误导性。
活线大脑的特点：
- 终其一生都在物理层面改变其回路，没有终点
- 将硬件与软件融合为一个连续的谱系——不存在清晰的层级划分
- 同时在多个层面发生变化：突触权重、受体分布、神经元结构、生化级联反应以及epigenome
一个有用的类比是步调层次（pace layers，最初由 Stewart Brand 用于描述城市）：时尚变化最快，治理变化较慢，建筑变化更慢，自然变化最慢。大脑同样拥有在不同时间尺度上运作的变化层次——从快速的生化级联反应，到深度固化的长期记忆结构。
这解释了Ribot 定律（神经学最古老的规律之一）：旧记忆比新记忆更稳定，因为随着时间推移，信息被逐渐固化到系统更深的层次中。
生命周期中的可塑性窗口
- 7 岁以下的儿童可以接受整个大脑半球的外科切除术（hemispherectomy），并保留接近正常的功能，仅表现出轻微的跛行。
- 视觉皮层固化相对较快——视觉世界较为稳定，因此系统会较早锁定。
- 运动皮层和躯体感觉皮层保持较高的可塑性——身体会持续变化（生长、受伤、学会使用自行车或冲浪板等新工具），因此系统必须保持灵活。
- 可塑性并非随年龄简单终止；保持新奇感、社交参与和挑战的老年人能够维持显著更多的认知功能。
“白板”问题
大脑在出生时并非一块白板。它出生时已预先配线，具备：
- 将感觉数据路由至正确脑区的能力（眼睛→视觉皮层，耳朵→听觉皮层）
- 语言吸收能力——人类天生就能吸收周围环境中的任何语言
- 社会学习与文化传承能力
进化策略：构建一个半成品大脑，让它去吸收所处的环境，而非构建一个完全硬编码的系统。这就是为什么人类与鳄鱼不同，能够在世代之间和不同环境中发生显著变化。
感觉替代的”土豆头”理论
Eagleman 提出，大脑将外周感觉器官视为即插即用的设备——它无需为每种新的输入类型重新建立核心运作原则。
证据：
- 人工耳蜗：内耳中的电极产生数字信号而非生物信号；大脑学会将其解读为听觉。
- 视网膜植入体：电极网格插入视网膜后可产生视觉体验，尽管其信号对大脑的”母语”而言是陌生的。
- 在动物物种中：热感应坑、电感受器、磁场传感器——不同的外设，相同的核心学习原则。
这对brain-computer interfaces具有直接启示：只要数据有用且稳定，大脑就能适应全新的输入格式。
脑机接口：机遇与局限
Eagleman 对临床应用（帕金森病、癫痫、瘫痪）领域的 BCI 持谨慎乐观态度，但对广泛的消费者应用持怀疑态度，原因在于：
- 开颅手术存在真实的死亡和感染风险
- 大脑已经能够快速适应非侵入性界面（触摸屏、语音助手等）
- 健康人是否会为了有限的速度提升而选择手术，目前尚不明朗
他更感兴趣的是非侵入性方法，即在不开颅的情况下实现与大脑之间的信息传输。
瘫痪患者控制机械臂的案例展示了双向适应：工程师优化算法以读取运动皮层信号，与此同时，患者的大脑也在重新布线，以更有效地控制设备——尤其是在食物或紧迫需求作为奖励的情况下。
大脑中的内群体/外群体偏见
Eagleman 的实验室利用功能性磁共振成像（fMRI）进行了empathy研究：
- 参与者观看被标注了宗教身份（基督徒、犹太人、穆斯林、无神论者、山达基教徒、印度教徒）的手被注射器针头刺入的画面。
- 所有宗教群体的参与者，其疼痛矩阵对内群体的手的激活程度均更强——这是一种低层级的神经学反应，而非单纯的认知判断。
- 当参与者通过抛硬币被随机分配到虚构的部落（“奥古斯丁派”与”查士丁尼派”）时，他们很快就表现出相同的内群体偏见，由此可见大脑进入部落状态是何等容易且随意。
启示：历史上被归因于意识形态或文化的行为，部分是由围绕群体归属的深层神经结构驱动的。
法律体系与神经科学
Eagleman 运营着一家名为科学与法律中心（Center for Science and Law）的非营利机构，致力于将神经科学融入刑事司法。
核心论点：
- 现行体系对截然不同的大脑状态（精神分裂症、精神病态、成瘾）适用千篇一律的量刑方式
- 目标并非消除问责制，而是将焦点从归咎转向何种康复策略最为适当
- 由具备专业知识的法官和陪审员组成的专业法庭（心理健康法庭、毒品法庭）展现出更好的结果
- 这些改革往往发生在地方财政无力承担大规模监禁的情况下
大脑能为 AI 提供哪些启示
Eagleman 认为，当前的人工神经网络缺失了生物智能的几个基本特征：
| 特征 | 人类大脑 | 当前 AI（如 GPT-3） |
|---|---|---|
| 相关性过滤 | 强——编码对生存和目标重要的信息 | 弱——处理所有输入，无内在优先级 |
| 他心模型 | 丰富——追踪特定他人的知识、欲望、恐惧 | 缺失——不具备心智理论 |
| 具身生存驱动 | 存在——死亡、饥饿、欲望 | 缺失 |
| 身体图式的实时适应 | 是——狼会咬断被夹住的腿 | 否——火星探测器 Curiosity 在一个车轮卡住时便陷入失败 |
他提出：“如果你想造机器人，就从胃开始” —— 意思是，任何真正具有适应性的系统，在能够以灵活、通用的形式涌现出智能之前，都需要内在的驱动力（饥饿、生存、相关性）。
单纯增加参数无法弥合这一鸿沟。架构必须改变，以纳入类似于”关心结果”的机制。
涉及概念
- neuroplasticity
- livewired brain
- hemispherectomy
- epigenetics
English Original 英文原文
The Livewired Brain: Neuroplasticity, Adaptation, and the Future of Intelligence
Summary
Neuroscientist David Eagleman discusses his book Livewired, arguing that the brain is fundamentally different from conventional hardware/software systems — it physically reconfigures itself throughout life in response to experience. The conversation covers the principles of neuroplasticity, the limits of brain-computer interfaces, the nature of intelligence, and what biological brains might teach us about building better AI systems.
Key Takeaways
- The brain is “livewired,” not merely plastic — it never stops rewiring itself; there is no final state, unlike a molded plastic object that holds its shape permanently.
- Different brain regions have different plasticity windows — the visual cortex hardens relatively quickly; the motor and somatosensory cortices remain more malleable because the body keeps changing.
- Plasticity diminishes with age partly due to motivation, not just biology — older people who remain cognitively and socially active can maintain significant neuroplasticity, even in the presence of physical Alzheimer’s pathology.
- Novelty and challenge are the primary drivers of brain change — routines reduce plasticity; being forced out of habitual patterns (as in 2020) actively promotes rewiring.
- The brain is a relevance-filtering machine — it encodes what matters for survival and personal goals, not all incoming data equally.
- The brain speaks any sensory dialect — cochlear implants and retinal implants work not because they replicate biological signals perfectly, but because the brain figures out how to interpret whatever data arrives.
- In-group/out-group bias is a low-level neurological response, not purely a cognitive or cultural one — it can be triggered by arbitrary labels within minutes of assignment.
- “Just-in-time” learning is superior to rote memorization — curiosity-driven, contextual learning produces stronger neural encoding than abstract, decontextualized education.
- For AI to approach human-level intelligence, it likely needs a survival drive and a sense of relevance, not just more parameters or larger training sets.
Detailed Notes
What Is “Livewired”?
Eagleman coined the term livewired to replace the traditional neuroplasticity framing. The word plastic (coined by William James over 100 years ago) implies a system that gets molded and then holds its shape — accurate for plastic manufacturing but misleading for the brain.
The livewired brain:
- Physically changes its circuitry throughout life, with no endpoint
- Blends hardware and software into a continuous spectrum — there are no clean layers
- Changes at multiple levels simultaneously: synaptic weights, receptor distribution, neuron structure, biochemical cascades, and the epigenome
A useful analogy is pace layers (originally from Stewart Brand describing cities): fashion changes fast, governance more slowly, buildings even more slowly, nature most slowly. The brain similarly has layers of change operating at different timescales — from rapid biochemical cascades down to deeply cemented long-term memory structures.
This explains Ribot’s Law (one of neurology’s oldest rules): older memories are more stable than newer ones because through time, information gets cemented into progressively deeper layers of the system.
Plasticity Windows Across the Lifespan
- Children under ~7 years can have an entire hemisphere surgically removed (hemispherectomy) and retain near-normal function, with only a slight limp.
- Visual cortex hardens relatively quickly — the visual world is stable, so the system locks in early.
- Motor and somatosensory cortices stay more malleable — bodies change (growth, injury, new tools like bicycles or surfboards), so the system must remain flexible.
- Plasticity doesn’t simply stop with age; older adults who maintain novelty, social engagement, and challenge preserve significantly more cognitive function.
The “Blank Slate” Question
The brain is not a blank slate at birth. It arrives pre-wired for:
- Routing sensory data to correct brain regions (eyes → visual cortex, ears → auditory cortex)
- Language absorption — humans are pre-configured to absorb whatever language surrounds them
- Social learning and cultural transmission
The evolutionary strategy: build a half-finished brain that absorbs its environment, rather than a fully hard-coded system. This is why humans, unlike alligators, change dramatically across generations and environments.
The “Potato Head” Theory of Sensory Substitution
Eagleman proposes that the brain treats peripheral sensory organs as plug-and-play devices — it doesn’t need to reinvent its core operating principles for each new input type.
Evidence:
- Cochlear implants: electrodes in the inner ear produce digital signals, not biological ones; the brain learns to interpret them as hearing.
- Retinal implants: electrode grids plugged into the retina allow visual experience even though the signal is foreign to the brain’s native language.
- Across animal species: heat pits, electroreceptors, magnetic field sensors — different peripherals, same core learning principle.
This has direct implications for brain-computer interfaces: the brain will adapt to novel input formats if the data is useful and consistent.
Brain-Computer Interfaces: Opportunity and Limitations
Eagleman is cautiously optimistic about BCI for clinical applications (Parkinson’s, epilepsy, paralysis) but skeptical of widespread consumer adoption because:
- Open-skull surgery carries real risks of death and infection
- The brain already adapts rapidly to non-invasive interfaces (touchscreens, voice assistants, etc.)
- It is unclear how many healthy people would elect surgery for modest speed improvements
He is more interested in non-invasive methods of getting information in and out of the brain without breaching the skull.
Paralyzed patients controlling robotic arms demonstrate bidirectional adaptation: engineers optimize algorithms to read motor cortex signals, while simultaneously the patient’s brain rewires itself to more effectively control the device — especially when food or urgent need is the reward.
In-Group/Out-Group Bias in the Brain
Eagleman’s lab conducted empathy studies using fMRI:
- Participants viewed hands labeled with religious identities (Christian, Jewish, Muslim, Atheist, Scientologist, Hindu) being stabbed with a syringe needle.
- The pain matrix activated more strongly for in-group hands across all religious groups — a low-level neurological response, not merely a cognitive judgment.
- When participants were randomly assigned to invented tribes (“Augustinian” vs. “Justinian”) via a coin toss, they rapidly showed the same in-group bias, demonstrating how easily and arbitrarily the brain enters tribal states.
Implication: behaviors historically attributed to ideology or culture are partly driven by deep neurological architecture around group membership.
The Legal System and Neuroscience
Eagleman runs a nonprofit called the Center for Science and Law, focused on integrating neuroscience into criminal justice.
Key argument:
- Current systems apply one-size-fits-all sentencing despite dramatically different underlying brain states (schizophrenia, psychopathy, addiction)
- The goal is not to remove accountability, but to shift from blame to what rehabilitation strategy is most appropriate
- Specialized courts (mental health courts, drug courts) staffed by judges and jurors with domain expertise show better outcomes
- These reforms tend to happen when counties run out of money and can no longer afford mass incarceration
What the Brain Can Teach AI
Eagleman argues current artificial neural networks are missing several fundamental features of biological intelligence:
| Feature | Human Brain | Current AI (e.g., GPT-3) |
|---|---|---|
| Relevance filtering | Strong — encodes what matters for survival and goals | Weak — processes all inputs without intrinsic prioritization |
| Model of other minds | Rich — tracks what specific people know, want, fear | Absent — no theory of mind |
| Embodied survival drive | Present — mortality, hunger, desire | Absent |
| Live adaptation of body schema | Yes — wolf chewing off trapped leg | No — Mars Rover Curiosity failed when one wheel got stuck |
He proposes: “If you want to build a robot, start with the stomach” — meaning any truly adaptive system needs intrinsic drives (hunger, survival, relevance) before intelligence can emerge in a flexible, generalizable form.
More parameters alone will not bridge this gap. The architecture must change to incorporate something analogous to caring about outcomes.
Mentioned Concepts
- neuroplasticity
- livewired brain
- hemispherectomy
- epigenetics