2 Neuroscience2.3 Basic techniques in neuroscience
2.3.5 Computational modelling and data analytic descriptions
1 min read1 figure
Written by Resident Doctor Committee · Published
Models and data analysis
A computational model states a hypothesis as mathematics or an algorithm. Data analysis finds patterns or predicts outcomes without explaining their cause.
Model type | Question answered |
|---|---|
Descriptive | What pattern is present? |
Mechanistic | What process could generate it? |
Normative | What is optimal under stated goals and constraints? |
A good fit does not prove a model uniquely correct.
Bayesian predictive processing
Predictive processing treats perception as Bayesian inference: the brain combines expectations with sensory evidence.
Term | Meaning |
|---|---|
Prior | The expectation held before sensory evidence arrives |
Likelihood | The evidence: how probable the input is under a possible state |
Posterior | The percept produced by combining prior and likelihood |
Prediction error | The mismatch signal between expectation and evidence |
Precision | The weighting each signal receives |
Behaviour that fits a Bayesian model does not show the brain calculating explicitly.
Other learning and decision models
Reinforcement learning updates value by prediction error: new value = old value + learning rate multiplied by prediction error.
Evidence accumulation models a decision as noisy evidence building to a boundary, set by drift rate, boundary separation, starting point and non-decision time.
Fitting and interpreting
Overfitting happens when spare flexibility lets a model fit noise as well as signal, and held-out data and complexity penalties check for it.
Cross-validation must match the intended use, so splitting trials from the same participants does not test performance in new patients.
Parameter recovery tests identifiability by simulating known values and fitting them back.
A fitted learning rate is not a direct measure of dopamine or impulsivity without separate validation.