哈佛大学Carlos R. Ponce小组宣布他们揭示了神经元调谐在视觉层次上动态地与对象和纹理流形对齐。2026年3月10日，国际知名学术期刊《自然—神经科学》发表了这一成果。

为了了解神经元优先考虑哪些特征，研究组对生成模型（从学习潜在空间合成新图像的深度网络）进行了主题化，允许V1、V4和后颞下皮层（PIT）的神经元通过闭环优化指导图像合成。该课题组人员比较了强调纹理的模型和强调物体结构的模型。尽管V1和V4与基于纹理的空间的一致性更强，但许多PIT神经元对这两种类型的优化图像的反应同样好，揭示了对共享的局部主题而不是整个对象模板的关注，并且这种对对象的一致性在他们的反应中出现得较晚。这些发现揭示了腹侧流的编码原理，并澄清了当前视觉模型的局限性。

据悉，视觉神经元对从纹理到物体的大量图像做出反应，但连接这些反应的规则尚不清楚。虽然对简单特征的调整在初级视觉皮层中已经很好地建立起来，但这个框架在高级区域被打破了，那里的神经元编码了多种和不可预测的特征。

附：英文原文

Title: Neuronal tuning aligns dynamically with object and texture manifolds across the visual hierarchy

Author: Wang, Binxu, Ponce, Carlos R.

Issue&Volume: 2026-03-10

Abstract: Visual neurons respond to a vast range of images, from textures to objects, but the rules linking these responses remain unclear. Although tuning to simple features is well established in the primary visual cortex, this framework breaks down in higher areas, where neurons encode diverse and unpredictable features. To ask what features neurons prioritize, we used generative models (deep networks that synthesize new images from a learned latent space), allowing neurons in V1, V4 and the posterior inferotemporal cortex (PIT) to guide image synthesis through closed-loop optimization. We compared models that emphasize texture versus those that emphasize object structure. Although V1 and V4 aligned more strongly with texture-based spaces, many PIT neurons responded equally well to both types of optimized images, revealing a focus on shared local motifs rather than whole-object templates, and this alignment to objects emerged later in their response. These findings reveal coding principles across the ventral stream and clarify the limits of current vision models.

DOI: 10.1038/s41593-026-02207-1

Source: https://www.nature.com/articles/s41593-026-02207-1