AI-generated videos to maximally drive a target brain region
296 points
• 7 days ago
• Article
Link
NEvo 是一个新型框架，通过合成高度优化的视频来揭示人脑特定视觉区域对何种视觉信息敏感。该系统不依赖静态图像或预先录制的视频片段，而是通过进化生成的 AI 视频来最大化激活所选的大脑目标区域。研究者借助大脑的"数字孪生"——一个训练用于预测视觉区域对不同输入反应的编码模型——系统地搜索能在任一皮层位置引发最大神经活动的精确视觉刺激。
合成过程由一种进化算法驱动，将视频的主体、光照、运动与氛围等属性视为种群中的"基因"。系统生成一批视频，利用编码模型打分，并迭代地保留、混合与变异表现最好的个体，直至预测激活达到最优。为降低计算开销，该过程分两阶段进行：先找到最有效的静帧，再进化出运动，将其变成约两秒钟的短片。
实验证明，这类合成视频在激发大脑活动方面明显优于人工设计的定位器和自然视频。更重要的是，动画剪辑始终强于其静止对应物，证明神经区域对动态信息高度敏感。将该方法应用于视觉皮层外侧流时，刺激呈现出清晰的功能演化：从 V1 的基本图案与运动，到前颞上沟的复杂社会互动。
通过在皮层表面进行探照灯映射，研究者能够观察视觉选择性如何向人物、面孔和社会动态偏移。即便从抽象元素（如简单堆叠的圆盘）出发，模型也能演化出独特且有意义的形态——在处理社会信息的区域出现类面孔结构，在运动敏感区域则呈现纯粹的运动模式。由此可得到一种精确且自动化的手段，用于分离并绘制视觉大脑各部分真正关注的特征。
NEvo is a novel framework designed to uncover what specific visual regions in the human brain respond to by synthesizing highly optimized videos. Instead of relying on static images or pre-recorded clips, the system works by evolving AI-generated videos that maximally activate a chosen target brain region. By leveraging a digital twin of the brain—an encoding model trained to predict how visual areas react to various inputs—the researchers can systematically search for the exact visual stimuli that trigger the highest level of neural activity in any given cortical location.
The synthesis process operates through an evolutionary algorithm that treats video attributes like subject, lighting, motion, and mood as genes within a population. The system generates a batch of videos, scores them using the encoding model, and iteratively keeps, mixes, and mutates the most effective versions until the predicted activation levels are optimized. To manage computational demands, the process happens in two stages, beginning with finding the most effective still image before evolving the motion to animate it into a short, two-second clip.
This approach demonstrates that these synthetic videos are significantly more effective at driving brain activity than either handcrafted localizers or natural videos. Crucially, the results confirm that neural regions are sensitive to dynamics, as the animated clips consistently outperform their frozen, static counterparts. When applied across the lateral stream of the visual cortex, the stimuli show a clear developmental arc, moving from basic patterns and motion in areas like V1 to complex, social interactions in the anterior superior temporal sulcus.
By using searchlight mapping across the cortical surface, researchers can observe how visual selectivity shifts toward people, faces, and social dynamics. Even when the optimization starts from abstract elements, like simple stacked discs, the model is able to conjure distinct, meaningful forms, such as face-like figures for social processing areas or pure movement patterns for motion-sensitive regions. This allows for a precise, automated way to isolate and map the specific features that each part of the visual brain truly cares about.
248 comments • Comments Link
一项新研究用人类大脑区域的 "digital twins" 来识别并生成能最大化刺激特定视觉神经通路的视频，目的在于更好地理解大脑功能并减少实验者偏差。
但滥用潜力巨大：该技术可能让平台创造出"surgically addictive"的内容，利用生物脆弱性劫持用户注意力，其效果远超过现有的 engagement algorithms 。
有人认为这只是 A/B testing 与上瘾式设计等现有营销手段的自然演进；另一些人则认为这是认知操纵方面的临界点，必须进行严格监管或制定行业范围的伦理准则。
目前对该模型效能存有怀疑，一些人指出生成的刺激并不理想，而且将 brain geography 映射到行为结果仍是一个重大且未证实的障碍。
人们常把这一情形类比于反乌托邦小说，比如 Infinite Jest 中的致命电影、 A Clockwork Orange 里的 "Ludovico Technique"，以及 Snow Crash 和 Echopraxia 等作品中描绘的心理条件化。
研发此类技术的专业环境正受到严格审视，批评者质疑那些开发出具大规模心理剥削潜力工具的研究者应承担的道德责任。
另一种观点认为，neuroscience research 本质中立，改进的大脑功能模型对推进 dementia 、 ADHD 与其他 mental health conditions 的临床治疗至关重要。
历史表明，针对 engagement 的技术优化通常跑赢监管，人们越来越担心利润驱动会不可避免地把这项研究推向商业化滥用的 "distractatron" 路径。
讨论中有一部分强调 human agency 的缺失：一旦大脑的机械触发点被完全映射，free will 的概念可能变得过时，社会将更易受到国家或企业的控制。
一些参与者指出，当前的 short-form content 已经在发挥类似 "mood organ" 的作用，能控制用户的情绪状态，即便没有先进的 generative models，整个行业也在朝这些目标迈进。
这场争论的核心是科学探索大脑 architecture 与被营利性实体利用该研究所带来的 existential risk 之间的紧张关系。支持者把这项工作视为理解 human cognition 与改善 medical outcomes 的必要步骤；批评者则将其定性为 "Torment Nexus"——一种为胁迫和成瘾创造了明确工具的技术。由此可见，对 tech industry 激励机制的深刻讽刺：在这些机制下，对 engagement 的渴求往往高于对 human well‑being 的关注，因此许多人主张在这些能力成熟之前，必须建立更健全的伦理框架。 • A new research project uses "digital twins" of human brain regions to identify and generate videos that maximally stimulate specific visual neural pathways, ostensibly to better understand brain function and reduce experimenter bias.
• The potential for misuse is profound, as this technology could allow platforms to create "surgically addictive" content that exploits biological vulnerabilities to hijack user attention far more effectively than current engagement algorithms.
• While some argue this is simply a natural evolution of existing marketing techniques like A/B testing and addictive design, others view it as an unprecedented threshold in cognitive manipulation that necessitates strict regulation or an industry-wide ethical code.
• Skepticism exists regarding the current efficacy of the model, with some noting that the generated stimuli are underwhelming and that mapping brain geography to behavioral outcomes remains a significant, unproven hurdle.
• Parallels are frequently drawn to dystopian fiction, including the lethal films in Infinite Jest, the "Ludovico Technique" of A Clockwork Orange, and the psychological conditioning described in works like Snow Crash and Echopraxia.
• The professional environment of those developing such technology is under intense scrutiny, with critics questioning the moral responsibility of researchers who produce tools with obvious potential for mass psychological exploitation.
• An alternative perspective suggests that neuroscience research is fundamentally neutral and that improved models of brain function are essential for advancing clinical treatments for dementia, ADHD, and other mental health conditions.
• History demonstrates that technological "optimizations" for engagement often outpace regulatory efforts, and there is a growing concern that profit motives will inevitably drive this research toward "distractatron"-style commercial abuse.
• A subset of the discussion emphasizes the lack of human agency, suggesting that once the mechanical triggers of the brain are fully mapped, the concept of free will may be rendered obsolete, leaving society vulnerable to state or corporate control.
• Some participants argue that current short-form content already functions as a "mood organ" capable of controlling user states, and that the industry is already moving toward these goals even without advanced generative models.
The discussion centers on the tension between scientific inquiry into brain architecture and the existential risk of its exploitation by profit-driven entities. While proponents characterize the work as a necessary step toward understanding human cognition and improving medical outcomes, critics identify it as a "Torment Nexus" event—a technology that creates clear tools for coercion and addiction. The dialogue reveals a deep-seated cynicism toward the tech industry's incentive structures, where the desire for engagement often supersedes human well-being, leading many to advocate for a more robust ethical framework before such capabilities reach maturity.