Face and Object Perception Notes
How does the brain get from "stripes and edges" (V1, mid-level vision) to "that's my mum's face / a cup"? It answers in three parts: where it happens (the "What" pathway), how objects are coded (templates vs structural descriptions), and what breaks (agnosia, prosopagnosia).
1. The "What" pathway: where recognition happens
Timeline of evidence
fMRI methods
Subtraction method: compare a condition with the process of interest against one without it. Example: intact face vs scrambled face. Your handwritten note ("same visual energy") is the key point: scrambled images keep the same low-level brightness and contrast, so any leftover activation is due to object structure, not simple light.
Op de Beeck et al. (2008): contrasts like faces > bodies + houses map out face-, body- and house-selective regions in a single person.
Decoding (Kay et al., 2008): train a model on voxel patterns for known images, then use the patterns to predict what a new image was. If you can predict it from a region, that region carries information about that object class.
Confusing bit: "analogous." Humans and monkeys aren't identical, but the human ventral occipito-temporal cortex does the same job as monkey IT. That's why monkey and human evidence are combined.
2. Why object recognition is a hard problem
It feels rapid, effortless and accurate, but the brain must handle:
Variety and familiarity: different cars and dinosaurs share a category.
Variable views: position, size, visible surfaces, lighting and shadows change constantly. Yet we keep object constancy.
Clutter: objects are rarely alone, so you need segmentation, figure/ground, grouping and filling-in (lecture 5 material).
Shape over colour or texture (Naor-Raz et al., 2003): a blue banana is recognised slower but still correctly, so colour helps but isn't critical. Conway (2018) adds that colour may carry information about behavioural relevance independent of identity (ripe vs rotten banana, healthy vs ill face). Your note says it well: we're still sensitive to colour, it isn't filtered out.
3. How are objects coded? ("Good Exam Q's")
The two camps
Evidence for each
For RBC: Biederman (1987) deleted contours in line drawings. Removing contours at concavities (which define how parts join) wrecks recognition ("Hard"). Removing mid-segments doesn't ("Easy"). So the relationships between parts matter.
For view-based:
Hayward & Tarr (1997): even with the same geons visible, error rate rises as rotation from the learned view increases (0 → 30°). RBC predicts a flat line (the blue "view-independent" line on the slide).
Logothetis et al. (1995): monkeys learned novel objects from specific views. Recognition was best for learned views and within ~40° of them, and they could interpolate between views. Single IT neurons respond differently to different views.
Problems for each theory
Templates: you'd need a lot of them (all letter fonts, all cow photos). View-based theories also struggle with the transformation step and with specifying what information is encoded. The Pasupathy & Connor curvature/polar model (lecture 5) offers a possible way to code these representations.
RBC: the view-dependence data above.
Current resolution: it's both. Type, complexity and familiarity matter:
Exemplar-specific ("my pet apatosaurus") → mostly view-based → Subordinate ("apatosaurus") → mixed → Entry-level ("dinosaur") → mostly structural description.
4. When it breaks: agnosia and prosopagnosia
Visual agnosia ("not knowing")
Impaired meaning of visually presented stimuli, despite intact low-level vision.
Objects can be recognised through touch or hearing, so knowledge of object properties is intact.
Not general intellectual decline or a language problem, and other categories (e.g. faces) can be spared.
Case: Kevin Chappell. Your notes say he can see colour and uses shapes/components to guess (long and silver = spoon or fork?).
Prosopagnosia (Bodamer, 1947): selective disruption of face perception, seen but not recognised as belonging to a particular person, including your own face. This is the "selective" loss that supports face specialisation.
Acquired: could recognise faces before a brain injury. Usually right occipital/temporal lobe damage.
Developmental/congenital: failed to develop face skills, with no known injury. May show subtle brain differences. It often goes undetected because people have no comparison, use compensatory strategies (voice, clothes, smell), and are sometimes misdiagnosed as behavioural or autistic.
Perception or memory?
Face perception is measured with tests like the CFPT (Cambridge Face Perception Test) and Glasgow Face Matching Test.
Face memory is measured with the CFMT (Cambridge Face Memory Test).
In the general population they're correlated but not identical (r = .528, n = 4265).
Dalrymple, Garrido & Duchaine (2014): a dissociation in adults, but not children. In children with DP, both perception and memory were impaired, but in adults only about half also had impaired face perception. So a person can fail face memory but pass face perception.
Liu et al. (2021): separate neural bases (face memory ↔ bilateral fusiform; face perception ↔ right posterior superior temporal sulcus) plus shared areas (right precuneus, right orbitofrontal cortex).
Why it matters: selective deficits reveal which processes are separable, which supports the idea of specialised systems.
5. Connections to earlier material
Lecture 5 (mid-level vision): segmentation, Gestalt grouping and curvature feed into shape coding. Curvature coding (Pasupathy & Connor) is proposed as a way to code view-based templates.
Lecture 1-4 (V1): contrast the "spots and stripes" simple cells with IT cells' huge receptive fields and complex selectivity, which is the hierarchy idea.
6. Likely exam questions
Compare and contrast structural description theory (Biederman) with view-based theories (Tarr). Which does the evidence favour? (Use the Hayward & Tarr, Logothetis, and Biederman 1987 evidence, then conclude: both, depending on the recognition level.)
What is a geon? Why are geons proposed to be viewpoint invariant?
Explain the subtraction method in fMRI. Why use scrambled images as the control?
What is object constancy, and why is it a challenge for recognition?
Describe the difference between acquired and developmental prosopagnosia. Why might developmental cases go undetected?
Is prosopagnosia a problem of perception or memory? Use tests and evidence.
How does the evidence from monkey IT support hierarchical models of perception?
7. Quick self-checks
Why can't Hayward & Tarr's rotation result be explained by pure RBC? (Same geons visible, yet performance still drops with rotation.)
A patient can describe a fork by touch but can't name it by sight. Agnosia or a language problem? (Agnosia, because touch recognition works.)
Someone scores normally on the CFPT but poorly on the CFMT. What does that suggest? (Face memory impairment with intact perception, i.e. a dissociation.)
Why is scrambling used rather than a blank screen? (It matches low-level visual energy.)
8. Memory hooks
Klüver-Bucy = "psychic blindness": sees but doesn't know.
Geons = an alphabet for objects; a cup and a bucket are the same "letters" in a different order.
Views vs parts: Tarr = photo album; Biederman = Lego kit.
The specificity slider: the more specific ("my pet"), the more view-based. The more general ("dinosaur"), the more structural.
CFPT = Perception (matching); CFMT = Memory (learn then recognise).
Ventral = What ("V" for "Vision of identity"); Dorsal = Where
The brain builds increasingly complex shape representations along the ventral "What" pathway, then matches them to what it knows. Recognition is not one jump from edges to "cup" or "mum". Objects may be stored as parts plus relations (Biederman) and as learned views (Tarr), and current thinking uses both depending on how specific the identification is. Agnosia and prosopagnosia show that seeing and identifying are separable.
Hook: Edges → shape and views → identity.
1. Where: the ventral "What" pathway
2. Why recognition is hard
Variety and familiarity: Different cars or dinosaurs belong to one category.
Variable views: Position, size, visible surfaces and lighting change, yet object constancy holds.
Clutter: Segmentation, figure/ground and grouping (lecture 5) must come first.
Shape vs colour: Shape is key. A blue banana is recognised slower but still correctly (Naor-Raz et al., 2003). Colour can signal state, such as ripeness or illness (Conway, 2018).
3. How: parts vs views ("Good Exam Q's")
4. What breaks
Visual agnosia: Impaired meaning from sight despite intact low-level vision. Touch and sound recognition work. It is not a language or general intellectual problem, and faces can be spared (Kevin Chappell).
Prosopagnosia (Bodamer, 1947): Selective loss of face identification, including your own face.
Acquired: Normal before injury, usually right occipito-temporal damage.
Developmental: Never developed, with no known injury. It is often undetected (no comparison), people use compensatory strategies (voice, clothes), and it can be misdiagnosed as behavioural or autistic tendencies.
Perception vs memory: CFPT (and Glasgow Face Matching Test) measure perception. CFMT measures memory. They correlate (r = .528, n = 4265) but are distinct. Dalrymple et al. (2014) found a dissociation in adults but not children. Liu et al. (2021) found separate and shared neural bases.
5. Sorting the material
6. Confusing parts, simplified
"Analogous to IT": Same job, not the same anatomy. That's why human recordings are used to confirm the monkey story.
Object-centred vs viewpoint-based: Object-centred means "cylinder with handle on the side", which survives rotation. Viewpoint-based means "how it looks from here".
Why Hayward & Tarr don't kill RBC: They show geons alone don't explain performance. Biederman's concavity data show parts still matter.
Agnosia vs other problems: Only the sight → meaning step fails.
Perception vs memory: Perception means judging whether two faces on screen match. Memory means learning a face and recognising it later.
FFA is "preferential": Faces > bodies and houses is a difference in degree. This is my caution, not the lecturer's, so don't present it as a slide claim.
7. Glossary (quick)
Ventral = What, dorsal = Where · IT = monkey end-of-pathway area · Receptive field = patch of visual field a cell responds to · Hierarchical model = simple → complex stages · Grandmother cell = one cell per thing (an idea, not proven) · Agnosia = not knowing what you see · Object constancy = same object despite changed view · Template = stored image matched to input · Structural description = parts + relations · Geon = a "letter" of 3D shape · Viewpoint invariant/dependent · Entry / subordinate / exemplar = dinosaur / apatosaurus / my pet apatosaurus · Prosopagnosia = selective face-identity loss · CFPT/CFMT = perception/memory tests · Dissociation = one ability impaired, another intact.
8. Study aids
Likely exam questions
Compare structural description and view-based theories: evidence for each, and current thinking?
What is a geon, and why might it support viewpoint invariance?
Explain the subtraction method. Why use scrambled images?
Why is recognition hard? Define object constancy.
Acquired vs developmental prosopagnosia: why might DP go undetected?
Perception or memory problem? Use the CFPT/CFMT and dissociation evidence.
Memory hooks: Tarr = photo album, Biederman = Lego kit · Specific → views, general → parts · Concavities = the glue between geons · Agnosia = sees, doesn't know · CFPT = Perception, CFMT = Memory.