Questions: Ventral Stream and Visual Object Recognition

Question 1

A neuron in inferotemporal (IT) cortex responds vigorously when shown a coffee mug viewed from the side. The mug is then rotated so it is viewed from above — a completely different pattern of pixels and edges. How would a typical IT neuron most likely respond?

• A. It would not respond — IT neurons are tuned to the specific pixel patterns of trained views
• B. It would respond vigorously — IT neurons are largely invariant to changes in viewing angle
• C. It would respond only if the rotation was gradual enough for the neuron to track the transformation
• D. It would respond only if the mug was also at the same retinal location as the original view

Answer: (B)

A defining property of the ventral stream is perceptual invariance: object identity is preserved in the neural representation despite changes in viewpoint, size, and illumination. While V1 neurons would respond completely differently to the two images (different edges, orientations, pixel patterns), IT neurons have receptive fields large enough to encompass much of the visual field and build representations that abstract away from low-level image properties. This is the core computational achievement of the ventral pathway.

Question 2

A patient with damage to the fusiform face area (FFA) struggles to identify familiar faces from photographs. According to the neuroscience of category-selective regions, which additional finding would you most expect?

• A. Complete inability to recognize any face, with perfectly normal recognition of all other objects
• B. Difficulty with fine-grained discrimination of other visually similar categories, especially in domains where the patient has expertise
• C. Normal face recognition in person but impaired performance on photographs, because the FFA requires 3D depth information
• D. Severely impaired recognition of all visual objects across all categories equally

Answer: (B)

The FFA is not a pure 'face detector' — it is better understood as a region specialized for fine-grained individuation within highly practiced visual categories. Prosopagnosia (face recognition deficit from FFA damage) often co-occurs with impaired discrimination within the patient's domain of visual expertise (e.g., car experts struggling to tell apart car models). A response to any and only faces would mean the region was truly face-exclusive, which it is not.

Question 3

In the ventral visual stream, neurons in earlier areas (V1, V2) have smaller receptive fields and respond to simpler features, while neurons in inferotemporal cortex have larger receptive fields and respond to complex shapes and objects.

• T. True
• F. False

Answer: True

This hierarchical organization is a well-established principle. V1 neurons respond to oriented edges in a small region of visual space. As you move anteriorly through the ventral stream, receptive fields progressively enlarge and feature preferences become increasingly complex — from contours and textures in V4 to whole objects and faces in IT. This progression implements the progressive abstraction needed for invariant object recognition.

Question 4

Category-selective regions like the fusiform face area (FFA) respond exclusively to their preferred category — a face-selective neuron produces no response at most to non-face objects.

• T. True
• F. False

Answer: False

This is a common misconception. The FFA is 'face-selective' in the sense that it responds more strongly to faces than other stimuli — not that it is silent to everything else. fMRI studies show the FFA responds to many object categories, just less strongly than to faces. Understanding it as a region with a strong prior toward face-like stimuli (holistic, fine-grained individuation) is more accurate than treating it as a binary face-detector switch.

Question 5

What computational problem does the ventral stream solve with perceptual invariance, and what is the key evidence that this is achieved through a hierarchical process rather than a single transformation?

Model answer: The problem is that an object's raw image (pixel values, edge patterns) changes dramatically with viewing angle, distance, illumination, and position, yet we recognize it as the same object. The ventral stream solves this by progressively transforming the representation through a hierarchy of areas, each pooling over more visual space and more feature combinations. Evidence includes the graded increase in receptive field size and feature complexity from V1 to IT, and the finding that earlier areas do not show invariance while IT neurons do.

The ventral stream's solution to invariance is fundamentally incremental — no single stage jumps from raw pixels to object identity. Instead, each area pools over the outputs of the previous area, progressively discarding information about low-level image statistics while preserving object identity. The hierarchical evidence (receptive field progression, feature tuning progression, and the fact that disrupting intermediate stages breaks recognition) is what distinguishes a true hierarchical computation from a single-step lookup table.