Cortex & brain

The visual cortex

VisualCortex models the ventral stream — the path from raw pixels to abstract object identity — using exactly one new primitive, the RetinotopicMap, and reusing the rest:

image [C,H,W]
  │  RetinotopicMap : tile into overlapping P×P receptive fields
  ▼
V1   : one weight-shared cortical-column PCN applied at every position  (= convolution)
  │  concatenate over the grid → V1 feature map
  ▼
V2…IT: HierarchicalNetwork (deep modular hierarchy) → abstract object latent
  ▼
readout (optional): PCN classifier on the IT latent

The retinotopic map extracts an image into overlapping patches and assembles them back overlap-averaged — that is the receptive-field machinery. Everything else is the standard engine: the shared V1 column autoencodes every patch (learning Gabor-like features, exactly as Rao & Ballard described), each higher area greedily encodes the one below, and the readout classifies by minimum free energy.

The cortical areas

The other regions are organised by system. Most are thin wrappers around the architectures of the previous page; the pure controllers (Thalamus, BasalGanglia, the neuromodulators) hold no persistent state and only emit signals.

Sensory

Area	What it is
AuditoryCortex	Tonotopic A1→belt→parabelt: the visual hierarchy applied to a 1-channel spectrogram.
SomatosensoryCortex	Somatotopic S1: the same microcircuit on a skin map.
DorsalStream	The "where/how" path: a per-frame visual cortex plus a deep recurrent network for motion.
RetinaLgn	Center-surround (difference-of-Gaussians) whitening front-end.
OlfactoryBulb	Decorrelating gain control feeding the piriform cortex.
GustatoryCortex	Taste content-addressable memory (a Hopfield attractor).
VestibularSystem	Self-motion integration into a head-direction code; partner of the grid cells.
FrontalEyeFields	Saccade selection: saliency plus inhibition-of-return over the retinotopic map.

Memory — the medial temporal lobe

Area	What it is
HippocampalFormation	Fast episodic store (Hopfield) plus offline replay consolidating into a neocortex (McClelland 1995).
EntorhinalCortex + GridCellEncoder	A hexagonal grid-cell code for 2-D position, plus a predictive spatial association.
PiriformCortex	Olfactory content-addressable memory.
PerirhinalCortex	Item familiarity and novelty: recognition memory.
ParahippocampalCortex	Scene and context gist — the "place area".

Executive and motor

Area	What it is
PrefrontalCortex	Gated working memory plus a hierarchical mixture of experts for task selection.
MotorCortex	An active-inference control loop that reaches for goals.
CerebellarCortex	A fast supervised forward model: given state and command, predict the next state.
PremotorCortex / SMA	Action sequencing: a temporal mixture of experts predicts the next action and feeds the motor cortex.

Integration and control

Area	What it is
AssociationCortex	Multimodal binding across the senses.
Thalamus	Precision and gain controller: allocates gain across areas by inverse free energy.
BasalGanglia	Action selection by energy gating — go/no-go.
DefaultModeNetwork	Offline generative imagination via sampling from the generative prior.
PosteriorParietalCortex	Coordinate transforms and intentional maps — the dorsal-to-motor bridge.
Claustrum	Global binding: a consensus broadcast (mean latent) and workspace coherence.

Neuromodulation

System	What it does
DopaminergicSystem	VTA/SNc dopamine: reward-prediction error with an average-reward baseline; modulates plasticity.
LocusCoeruleus	Noradrenaline: a global-surprise monitor whose phasic burst applies a broad precision reset across areas.
NucleusBasalis	Acetylcholine: per-area expected uncertainty that modulates each area's learning rate.
RapheNuclei	Serotonin: time horizon and patience, translated into the temporal-discount factor.

Limbic and affective

Area	What it is
Amygdala	Affective salience and valence; emits a salience gain that steers the thalamus.
OrbitofrontalCortex	Expected-value model feeding the basal ganglia.
CingulateCortex	Conflict and surprise monitor over area free energies — calls for top-down control.
InsularCortex	Interoception: integrates the body-state stream into a feeling latent.
Hypothalamus	Homeostatic set-points → drives, and the goal observation for the motor cortex.

Language

Area	What it is
WernickeArea	Comprehension: maps a token sequence to a semantic latent.
BrocaArea	Production: decodes a semantic latent into an utterance, paired with the motor cortex.
AngularGyrus	A relational semantic hub: completes a concept from its related concepts on a graph.

A Thousand Brains system

PredictiveCoding.Brain implements the Hawkins/Numenta papers (P1–P9, from A Thousand Brains to the 2025 sensorimotor-intelligence work) as a hard rule: a task stays on the roadmap only if it traces back to one of them. The end state — shipped, with the roadmap closed — is many cortical columns that each learn a model of an object in their own reference frames and vote to reach consensus:

• Reference frames — grid-cell-style pose encoding in 2-D and full 3-D (SE(3), quaternion algebra), with motor-driven path integration closing the loop.
• Voting — columns exchange evidence and converge on the object and pose consistent with all of them, so recognition works from partial, moving observation.
• Pose invariance — recognition from pairwise displacement signatures survives arbitrary rigid motion and occlusion, in 2-D and 3-D.
• Real sensors — a movable patch sensor scans real images; the same sensorimotor loop runs unchanged on camera frames.
• Cross-modal voting — columns over different modalities agree on the same object.

A benchmark and demo ship with it (demo option 9), rendering an animated, self-contained web page with the episode player, the benchmark and a real-sensor scan; 109 unit tests cover the stack.

The desktop demo

A WinUI 3 app trains and visualises several of these models, one page per example running on a background thread: the visual cortex classifying and reconstructing noisy shapes, Hopfield pattern completion, entorhinal grid-cell firing maps, a spiking network, and an active-inference agent learning a world model and reaching a goal.