Active Inference, The Learn Arc — Part 31: Session §5.4 — The Brain Map

Series: The Learn Arc — 50 posts teaching Active Inference through a live BEAM-native workbench. ← Part 30: Session 5.3. This is Part 31.

The session

Chapter 5, §4. Session title: Brain map. Route: /learn/session/5/s4_brain_map.

Chapter 5's final session folds Sessions 5.1–5.3 into one labeled diagram: a brain map where each cortical area is annotated with which factor-graph role it plays and which neuromodulator sets its precision. It's the most committed empirical claim in the book.

The map in one paragraph

Sensory cortex: bottom of the hierarchy. Likelihood factors; ACh modulates precision on observation errors. Primary sensory areas (V1, S1, A1) feed prediction errors upward. Association cortex: intermediate layers that infer higher-level latents (object identity, scene category) from sensory residuals. Prefrontal cortex: top of the hierarchy. Carries the slowest-changing factors — policy selection (Eq. 4.14 softmax, DA-precision-gated) and long-horizon preference evaluation (5-HT-gated). Hippocampus: the episodic-memory factor that grounds the generative model in past trajectories. Basal ganglia: the action-selection substrate that implements the softmax.

One sentence per area. Each sentence is a prediction. The combined picture is the strongest version of the theory.

Why this map matters for the Workbench

A lot of the Workbench's debugging surfaces have direct analogues on the brain map:

• /equations — the list of factors the brain hypothetically implements.
• /glass/agent/<id> — a per-agent factor-graph trace. The brain's equivalent would be an fMRI-plus-MEG record of every cortical area firing; the agent is cheaper to audit.
• /models — the taxonomy of model families. Each family corresponds to a different brain system (POMDPs ≈ sensorimotor, hierarchical ≈ cortical hierarchy, continuous-time ≈ motor reflex).

Session 5.4 is where you realise the Workbench's architecture is not accidental. It mirrors the brain map Chapter 5 draws, one surface at a time.

The four test predictions

Chapter 5 closes with four specific falsifiable predictions the brain map licenses:

1. Sensory-area lesions should look like low-observation-precision behavior. Patients with primary sensory lesions should over-rely on priors, not just lose sensation.
2. Prefrontal lesions should look like low-policy-precision behavior. Impaired policy softmax → indecision, perseveration. Documented in the PFC-damage literature.
3. Basal-ganglia disruption should disrupt action selection specifically, not belief-update. A patient with basal-ganglia damage should still update beliefs normally but fail to pick actions. Documented in Parkinson's.
4. Hippocampal lesions should disrupt prior over D. Patients with hippocampal damage should fail to condition beliefs on past episodes. Documented in amnesia.

All four match major neurological syndromes. The match is not proof of the theory — but it's striking.

The honest limitations

Session 5.4 ends with a clear-eyed list of what the brain map doesn't predict yet:

• Consciousness — not addressed at all.
• Specific empirical binding problems (how does one area know which signals go together?).
• Emotion, in detail — gestured at (interoception → affect) but not quantitative.
• Sleep, dreaming, default-mode dynamics.

All four are active research areas. The book is straightforward about "we don't have this yet." That honesty is why Session 5.4 is worth taking seriously.

The concepts this session surfaces

• Brain map — cortical areas + neuromodulators + factor-graph roles.
• Hierarchical cortex — sensory → association → prefrontal.
• Clinical prediction — behavioral signature you'd expect from a lesion.
• Open problem — what the theory doesn't yet explain.

The quiz

Q: The book's brain-map claim predicts that damage to prefrontal cortex should produce:

• ☐ Hallucinations.
• ☐ Inability to pick between plans; perseveration. ✓
• ☐ Aphasia.
• ☐ Complete loss of sensory experience.

Why: PFC is the top of the hierarchy in the brain map; it carries the policy softmax (Eq. 4.14), gated by DA. Damage → flat or stuck policy posterior → difficulty selecting actions, perseveration. Aphasia is a language-cortex issue; hallucinations are low-ACh; sensory loss is primary-sensory-area damage.

Run it yourself

• /learn/session/5/s4_brain_map — session page.
• /models — model-family taxonomy that mirrors the brain regions.
• /guide/jido — the Jido knowledgebase, where agent architecture is discussed in depth.
• /cookbook/hierarchical-context-switch — hierarchical composition, the Workbench analog of cortical hierarchy.

The mental move

Chapter 5 has given you a falsifiable theory of how brains implement Active Inference. You may or may not believe it. What matters for the rest of the series is that every neuroscience claim the book makes from here on will trace back to this map. Keep the diagram handy. Chapters 9 and 10 reference it heavily.

Next

Part 32: Session §6.1 — States, observations, actions. Chapter 6's opening session. The design template, in its first session. Back to the three lists from Session 4.1, but this time as a practice — working through two extended examples end to end.

---

⭐ Repo: github.com/TMDLRG/TheORCHESTRATEActiveInferenceWorkbench · MIT license

📖 Active Inference, Parr, Pezzulo, Friston — MIT Press 2022, CC BY-NC-ND: mitpress.mit.edu/9780262045353/active-inference

← Part 30: Session 5.3 · Part 31: Session 5.4 (this post) · Part 32: Session 6.1 → coming soon