Neural Computation, Representations, and Theories of Brain Coding
From Single Units to Populations
Population Codes
Modern views of neural coding emphasize populations rather than isolated neurons. If denotes the response of neuron , then information may be distributed across a vector , with computation emerging from geometry in population activity space. This motivates low-dimensional latent dynamical models, manifold learning, and subspace methods.
Core Theoretical Ideas
Dynamical Systems View
State-Space Modeling
A common formalism writes latent neural state evolution as and observations as Here may represent latent cognitive, motor, or neural variables; may represent stimuli or interventions; and is the measured signal.
Why Latent State Models Matter
They can disentangle meaningful cognitive dynamics from noisy measurements and provide a bridge between neuroscience theory and machine learning.
Competing Explanatory Lenses
| Lens | What it explains best | Main limitation |
|---|---|---|
| Encoding/decoding | Information mapping | Can miss internal dynamics |
| Dynamical systems | Time evolution and trajectories | May be hard to interpret causally |
| Normative theories | What optimal computation might be | Often underconstrained |
| Circuit models | Mechanistic microstructure | Can be difficult to fit at scale |
What is a key advantage of modeling neural activity in a latent state space?
Latent models can uncover structured, lower-dimensional trajectories that explain recorded activity.
Correct answer: It can reveal compact dynamics underlying high-dimensional observations
Write the difference between an observation model and a latent transition model.
This is the standard state-space decomposition.
Correct answer: The observation model maps latent variables to measured signals, while the latent transition model describes how the hidden state evolves over time.
Which claim is strongest regarding predictive coding?
Predictive coding is a family of theories in which the brain is seen as actively predicting sensory input.
Correct answer: Brains may minimize prediction error by comparing top-down expectations and bottom-up input