Ventral Stream Object Recognition and CNN Models

Object recognition depends on central vision and foveation, allowing rapid identification despite changes in viewpoint, illumination, occlusion, and non‑rigid deformation. Feed‑forward processing dominates the ventral stream, with visual signals reaching V1 in about 60 ms and IT in roughly 100 ms. IT neurons exhibit invariance to position and size, so that object identity becomes explicit early in the visual hierarchy.

Computational Basis of Recognition

Population coding treats the responses of many neurons as points on a high‑dimensional manifold. Linear classifiers project these points onto a vector perpendicular to a separating hyperplane, revealing that IT responses are largely linearly separable for object categories. Recordings from IT sites confirm that simple linear read‑outs can decode object identity with high accuracy.

Specialization for Object Classes

Face‑selective patches form segregated regions that respond preferentially to faces. Behavioral phenomena such as inversion effects, the Thatcher illusion, and prosopagnosia illustrate the modular nature of face processing. Disrupting face ability while leaving other visual functions intact provides strong evidence for functional specialization in the brain.

Representational Dissimilarity Matrices

A representational dissimilarity matrix (RDM) records 1 – correlation between neural responses to every pair of images, visualizing the geometry of the representation. Block‑diagonal structures in RDMs reveal category selectivity (e.g., animate versus inanimate). By constructing RDMs for humans, monkeys, and computational models, researchers can quantitatively compare visual systems across species.

Artificial Neural Networks as Models

Convolutional neural networks (CNNs) implement cascades of filtering, pooling, and normalization, learned through gradient descent. Intermediate CNN layers best predict activity in V4, while deeper layers align with IT responses, mirroring the hierarchical organization of the ventral stream. Encoding models fit a linear mapping from CNN unit activations to brain voxel responses, enabling precise prediction of neural activity.

Texture Recognition

Texture is defined by the average statistical properties of image regions. Energy‑based models compute squared, averaged filter responses to predict texture boundaries, while texture synthesis iteratively matches histograms of filter subbands to a target texture. Successful synthesis—where synthetic textures look like the original—tests whether a representation captures human texture perception.

Mechanisms in Practice

• Linear classifier – projects neural response vectors onto a separating hyperplane, providing a simple decision rule for object categories.
• Encoding model – learns a linear mapping from model features (e.g., CNN activations) to voxel responses, allowing prediction of brain activity from artificial representations.
• Texture synthesis – uses pyramid representations and histogram matching of subband statistics to generate images that share the same texture statistics as a target, probing the adequacy of texture representations.