Face-Space Coding & Prosopagnosia Simulator

Fusiform Population

Lesion severity 0%

Neural noise σ_n 0.05

Face-space distinctiveness 2.0

Trial Readout

Click “Present New Face” to begin.

Accuracy

—

Confidence

—

Trials

Active units

48/48

How it works

Faces are modelled as points in a 3D norm-based face-space (Valentine, 1991): each identity is a vector of feature deviations from an average face, and perceptual similarity is Euclidean distance. A population of 48 face-selective units, standing in for fusiform face area (FFA) neurons, each has a preferred location c_i in this space and a Gaussian tuning curve:

r_i = exp( -|x − c_i|² / (2σ²) ) + noise(σ_n)

When a face is viewed, its true position is corrupted by sensory noise to give a probe x. The population response is decoded back into face-space by a weighted centroid — the same population-vector scheme used for direction coding in motor cortex:

x̂ = Σ r_i · c_i  /  Σ r_i

The nearest stored identity to x̂ is the model's "recognition". Lesion severity silences a fraction of units (r_i forced to 0) — a stand-in for the fusiform damage seen in acquired prosopagnosia — which coarsens the decode and raises error rate, especially for identities close together in face-space (low distinctiveness), matching the clinical pattern in which prosopagnosic patients confuse similar-looking familiar faces while general object recognition stays intact.

Lesion severity — fraction of the population silenced, from 0 (healthy) to 90%.
Neural noise — trial-to-trial response noise added to every active unit.
Distinctiveness — radius of the 8 stored identities from the space's centre; low values crowd faces together.
Present New Face — draws a random identity, samples a noisy probe, and decodes it live.

Stored identity Probe (noisy input) Decoded estimate

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.