Robots Atlas>ROBOTS ATLAS
Evaluation

Neural Predictivity

2018ActivePublished: 29 September 2026Updated: 29 September 2026Published
Key innovation
A NeuroAI metric: how well a model activations predict measured neural (brain) responses — usually via linear regression (an encoding model).
Category
Evaluation
Abstraction level
Primitive
Operation level
Evaluation (runtime)Model
Use cases
Rating models as "models of the brain" (Brain-Score)Selecting layers that best map cortexNeuroAI research on model–brain agreementAnalysing brain-guided methods (NARI/NARF)

How it works

For a shared stimulus set, one collects model activations and neural responses. An encoding model (ridge regression) is fit from activations to neural activity on part of the data, and on held-out data the correlation (e.g. Pearson) between prediction and measurement is computed. The result, often normalised by measurement noise (noise ceiling), is the neural predictivity of a layer/model.

Problem solved

One needs to quantitatively assess whether a model processes information like the brain. Neural predictivity gives a number: how well the model representations predict real neural activity.

Components

Encoding model (ridge regression)Prediction

A mapping from model activations to neural activity.

Cross-validation + noise ceilingScoring

Held-out evaluation, normalised by measurement noise.

Evolution

2014
Deep models as predictors of visual cortex (Yamins, Khaligh-Razavi)
Inflection point
2018
Brain-Score formalises neural predictivity as a benchmark
Inflection point