大脑的工作原理、治愈失明与神经工程 | Dr. E.J. Chichilnisky
摘要
斯坦福大学神经外科与眼科学教授 Dr. E.J. Chichilnisky 阐述了视网膜如何通过不同细胞类型对视觉信息进行编码，以及这一知识如何被应用于构建智能视网膜植入物——这种装置有望恢复乃至增强人类视觉。对话还涉及科学探索的哲学、细胞类型特异性在神经工程中的重要性，以及如何应对非线性的职业发展路径。
核心要点
- 视网膜可以说是整个大脑中被理解得最透彻的神经回路，使其成为神经假体和脑机接口研究的理想起点。
- 视网膜同时向大脑传送约20路并行的视觉”影片”，每一路通过不同细胞类型编码不同的视觉特征（颜色、运动、边缘等）。
- 现有视网膜植入物无法恢复高质量视觉，原因在于它们将视网膜视作简单的像素网格，忽略了不同细胞类型及其特定的神经编码。
- 下一代”智能”视网膜植入物将采用三步流程：记录 → 校准 → 刺激，以适应每位患者视网膜回路的独特”语言”。
- 人工智能与机器学习对于将复杂视觉场景转化为各细胞类型应产生的精确电活动模式至关重要。
- 同一套用于恢复视觉的装置也可用于增强视觉——使人获得超越正常人类能力的感知体验。
- 药物干预（如dopamine、血清素类药物）作用广泛且不精确；细胞类型特异性电刺激代表了目前可设想的最高精度的大脑调控方式。
- 理解视网膜细胞类型对于与其他脑区的接口直接相关，包括视觉皮层，乃至潜在的与记忆相关的海马体。
- 非线性的职业路径——包括更换研究生项目、花数年时间从事舞蹈——可以是发现真正方向的一种策略性方式，而非失败。
详细笔记
视觉的起点：视网膜
- 视觉始于retina（视网膜），这是位于眼球后部的一层薄薄的神经组织。
- 视网膜具有三个功能层：
- 感光细胞——将光转化为电信号（“像素探测器”）；高度特化，易于死亡（由此引发macular degeneration（黄斑变性）和retinitis pigmentosa（视网膜色素变性）等疾病）。
- 中间处理层——数十种细胞类型，从感光细胞的原始信号中提取视觉特征。
- Retinal ganglion cells（视网膜神经节细胞，RGCs）——输出层；约20种不同类型，通过视神经将处理后的视觉信息发送至大脑。
20路”影片”：视网膜神经节细胞类型
- 约20种 RGC 类型中的每一种都覆盖整个视觉场景，但各自提取不同的特征：
- 空间细节（精细结构）
- 运动（移动物体）
- 颜色/波长（不同颜色通道）
- 亮度增量与减量（on细胞与off细胞）
- 类比：如同20种不同的 Photoshop 滤镜或并行影片同时发送至大脑，再由大脑将其整合为统一的视觉体验。
- 7种细胞类型已被充分表征，约占所有 RGC 输出的70%。
- 约15种额外细胞类型近年才开始被识别；部分细胞显示出意想不到的”蜘蛛状”或多团块空间响应图谱，其功能尚不明确。
视网膜研究的开展方式
- 视网膜取自脑死亡器官捐献者，在心脏支持终止后数分钟内摘取。
- 将眼球半切，视网膜铺平，取3×3毫米的组织片置于512通道电极阵列（“钉床”）上。
- 两种实验模式：
- 记录：向活体视网膜投射受控光刺激；记录 RGC 本应发送至大脑的脉冲放电模式。
- 刺激：通过电极施加电流直接激活 RGC——这是设计假体装置的关键。
- 采用闪烁棋盘格（“电视雪花”）刺激作为无偏差方法，通过将脉冲与先前视觉图案进行反向相关，在约30分钟内同时表征数百个细胞。
现有视网膜植入物的局限
- 现有设备（如 Argus II）将视网膜视为均匀的像素网格——不具备细胞类型特异性。
- 结果：患者能感知到明亮的门廊轮廓或大块光斑，但无法感知空间细节、颜色或复杂物体。
- 根本问题：数十年的视网膜科学成果均未被整合进现有植入物。
- 类比：就像将一个管弦乐队的乐谱打乱——你或许能辨认出某段旋律，但听不到完整的交响乐。
智能视网膜植入物：三步架构
- 记录——植入视网膜上约2毫米的芯片，识别该特定患者视网膜中存在的细胞类型及其电信号特征。
- 校准——通过刺激和记录，精确绘制哪些电极以何种概率激活哪些细胞的图谱（构建个性化刺激表）。
- 刺激——利用对**neural code（神经编码）**的了解（即每种细胞类型应该对任意图像产生何种放电），以正确的序列和时序激活细胞，从而再现自然的视网膜输出。
- 人工智能与机器学习负责处理从视觉场景到细胞特异性放电模式的复杂转换，并实现持续自适应调整。
视觉增强与更广泛的神经工程
- 同一植入平台理论上不仅可用于恢复，还可：
- 将空间分辨率提升至超越正常人类的极限。
- 向独立的细胞类型通路同时传递并行视觉信息流（例如，通过侏儒细胞阅读文字的同时，通过伞细胞监测运动）。
- 开启未增强人类从未体验过的全新视觉感知。
- 视网膜被明确定位为与所有大脑回路接口的概念验证——在此习得的知识同样适用于视觉皮层、海马体及更多脑区。
- 与药理学相比：dopamine或血清素类药物作用于全脑受体（特异性低）；细胞类型靶向电刺激代表了神经调控中可实现的最高精度。
跨物种视觉系统的比较
- 人类拥有3种视锥感光细胞（红、绿、蓝），这也解释了为何电视只需三种原色。
- 螳螂虾：拥有16种以上感光细胞类型——在它们眼中，人类的色觉极为贫乏。
- 颊窝蝮蛇：将红外线/热辐射作为视觉系统的一部分加以感知。
- 啮齿动物：拥有专门探测逼近的暗影（掠食性鸟类）的 RGC 类型——对灵长类动物则不那么重要。
- 每个物种的视网膜都受进化压力塑造，以提取与其生态位相关的特征。
职业导航：通往卓越的非线性路径
- Dr. Chichilnisky 经历了三个不同的研究生项目才找到自己的研究方向。
- 曾花数年时间从事专业舞蹈——这是一个深思熟虑的选择，帮助他厘清了自己的价值观与方向。
- 核心洞见：漫游不是失败；探索是识别你真正热衷于解决哪类问题的合理且往往必要的方式。
- 找到一个你认为深度有意义的问题——一个你希望彻底理解到能用数学公式表达并设计出解决方案的问题——比遵循传统的线性路径更为重要。
相关概念
- retina
- retinal ganglion cells
- photoreceptors
- neural code
- macular degeneration
- retinitis pigmentosa
- retinal prosthesis
- neuroengineering
- brain-computer interface
- neural augmentation
- visual cortex
- hippocampus
- electrophysiology
- cell types
- dopamine
- serotonin
- machine learning in neuroscience
- circadian rhythm
English Original 英文原文
How the Brain Works, Curing Blindness & Neural Engineering | Dr. E.J. Chichilnisky
Summary
Dr. E.J. Chichilnisky, professor of neurosurgery and ophthalmology at Stanford, explains how the retina encodes visual information through distinct cell types and how that knowledge is being applied to build smart retinal implants that could restore—and potentially augment—human vision. The conversation also covers the philosophy of scientific exploration, the importance of cell-type specificity in neuroengineering, and navigating a non-linear career path.
Key Takeaways
- The retina is arguably the best-understood circuit in the entire brain, making it the ideal starting point for neural prosthetics and brain-computer interfaces.
- The retina sends ~20 parallel “movies” of the visual world to the brain simultaneously, each encoding a different feature (color, motion, edges, etc.) via distinct cell types.
- Current retinal implants fail to restore high-quality vision because they treat the retina like a simple pixel grid, ignoring the distinct cell types and their specific neural codes.
- A next-generation “smart” retinal implant would use a three-step process: record → calibrate → stimulate to speak the language of each individual patient’s retinal circuitry.
- AI and machine learning are essential for translating complex visual scenes into the precise patterns of electrical activity each cell type should produce.
- The same device designed to restore vision could also be used to augment vision—enabling perceptual experiences beyond normal human capability.
- Drug-based interventions (e.g., dopamine, serotonin agents) are broad and imprecise; cell-type-specific electrical stimulation represents the highest precision level of brain modulation currently conceivable.
- Understanding the retina’s cell types has direct implications for interfacing with other brain regions, including the visual cortex and potentially the hippocampus for memory.
- A non-linear career path—including switching graduate programs and taking years off to pursue dance—can be a strategic way to discover one’s true direction, not a failure.
Detailed Notes
How Vision Begins: The Retina
- Vision starts in the retina, a thin sheet of neural tissue at the back of the eye.
- The retina has three functional layers:
- Photoreceptor cells – Convert light into electrical signals (“pixel detectors”); highly specialized and vulnerable to death (giving rise to conditions like macular degeneration and retinitis pigmentosa).
- Intermediate processing layer – Dozens of cell types that extract visual features from raw photoreceptor signals.
- Retinal ganglion cells (RGCs) – The output layer; ~20 distinct types send processed visual information via the optic nerve to the brain.
The 20 “Movies”: Retinal Ganglion Cell Types
- Each of the ~20 RGC types covers the entire visual scene but extracts a different feature:
- Spatial detail (fine structure)
- Motion (moving objects)
- Color/wavelength (different color channels)
- Luminance increments vs. decrements (on-cells vs. off-cells)
- Analogy: Like 20 different Photoshop filters or parallel movies sent simultaneously to the brain, which then assembles them into unified visual experience.
- 7 cell types are well-characterized and account for ~70% of all RGC output.
- ~15 additional cell types are only recently being identified; some show unexpected “spidery” or multi-blob spatial response profiles whose functions remain unknown.
How Retinal Research Is Conducted
- Retinas are obtained from brain-dead organ donors, harvested within minutes of cardiac support ending.
- The eye is hemisected, the retina laid flat, and a 3×3 mm piece is placed on a 512-channel electrode array (“bed of nails”).
- Two experimental modes:
- Recording: Shine controlled light stimuli onto the living retina; record the spike patterns RGCs would have sent to the brain.
- Stimulation: Pass current through electrodes to activate RGCs directly—key for designing prosthetic devices.
- A flickering checkerboard (“TV snow”) stimulus is used as an unbiased method to characterize hundreds of cells simultaneously in ~30 minutes by reverse-correlating spikes with preceding visual patterns.
Why Current Retinal Implants Fall Short
- Existing devices (e.g., Argus II) treat the retina as a uniform pixel grid—no cell-type specificity.
- Results: patients can detect bright doorways or large blobs of light but cannot perceive spatial detail, color, or complex objects.
- Fundamental problem: none of the decades of retinal science has been incorporated into existing implants.
- Analogy: Like scattering an orchestra’s sheet music—you might recognize a tune, but there is no coherent symphony.
The Smart Retinal Implant: Three-Step Architecture
- Record – A ~2mm chip implanted on the retina identifies which cell types are present and their electrical signatures in that specific patient’s retina.
- Calibrate – Stimulate and record to map precisely which electrodes activate which cells at what probability (building a personalized stimulation table).
- Stimulate – Using knowledge of the neural code (what each cell type should fire for any given image), activate cells in the correct sequence and timing to reproduce natural retinal output.
- AI and machine learning handle the complex transformation from visual scene to cell-specific spike patterns and enable continuous adaptation.
Vision Augmentation and Broader Neural Engineering
- The same implant platform used for restoration could theoretically:
- Augment spatial resolution beyond normal human limits.
- Deliver parallel visual information streams to independent cell-type pathways simultaneously (e.g., reading text via midget cells while monitoring motion via parasol cells).
- Enable novel visual sensations not experienced by unaugmented humans.
- The retina is explicitly positioned as the proof-of-concept for interfacing with all brain circuits—what is learned here applies to the visual cortex, hippocampus, and beyond.
- Comparison to pharmacology: drugs like dopamine or serotonin agents affect receptors brain-wide (low specificity); cell-type-targeted electrical stimulation represents the highest achievable precision in neural modulation.
Comparing Visual Systems Across Species
- Humans have 3 types of cone photoreceptors (red, green, blue), explaining why TVs only need three color primaries.
- Mantis shrimp: 16+ photoreceptor types—would find human color vision impoverished.
- Pit vipers: Detect infrared/heat as part of their visual system.
- Rodents: Have RGC types tuned to detect looming dark shadows (predatory birds)—less relevant to primates.
- Each species’ retina is shaped by evolutionary pressure to extract features relevant to its ecological niche.
Career Navigation: Non-Linear Paths to Excellence
- Dr. Chichilnisky moved through three different graduate programs before finding his focus.
- Took several years off to dance professionally—a deliberate choice that clarified his values and direction.
- Key insight: wandering is not failure; exploration is a legitimate and often necessary method for identifying what problems you are truly passionate about solving.
- Finding a problem you find deeply satisfying—one where you want to understand it so completely you can write it as a mathematical formula and engineer a solution—is more important than following a conventional linear path.
Mentioned Concepts
- retina
- retinal ganglion cells
- photoreceptors
- neural code
- macular degeneration
- retinitis pigmentosa
- retinal prosthesis
- neuroengineering
- brain-computer interface
- neural augmentation
- visual cortex
- hippocampus
- electrophysiology
- cell types
- dopamine
- serotonin
- machine learning in neuroscience
- circadian rhythm