Active Inference, The Learn Arc — Part 30: Session §5.3 — Neuromodulation, the Precision Knobs

#activeinference #neuroscience #ai #elixir

Series: The Learn Arc — 50 posts teaching Active Inference through a live BEAM-native workbench. ← Part 29: Session 5.2. This is Part 30.

The session

Chapter 5, §3. Session title: Neuromodulation. Route: /learn/session/5/s3_neuromodulation.

Session 5.2 argued that prediction errors are weighted by precisions. Session 5.3 is the payoff: each of the four major neuromodulators in the brain is claimed to be the physical implementation of one specific precision. If the theory is right, this is the clearest mapping from a mathematical parameter to a biological substance the literature has ever produced.

The table

Neuromodulator	Function in the brain	In the math
Acetylcholine (ACh)	Sensory gain, attention	Precision on the observation (likelihood) term
Noradrenaline (NA)	Arousal, novelty, surprise	Precision on temporal transition terms
Dopamine (DA)	Policy selection, salience	Precision on the policy softmax (γ in Eq. 4.14)
Serotonin (5-HT)	Patience, long-horizon valuation	Precision on preference (C vector)

Four substances. Four distinct roles. Every one is a specific, testable claim about what happens when you deplete or enhance that substance.

The clinical signatures

The testability is real. Each prediction the table makes maps to a known clinical syndrome:

Low ACh → over-reliance on priors, under-weighting of sensory evidence. Seen in: Alzheimer's disease, where cholinergic loss is profound. Behavioral: hallucinations, confabulation, failure to update beliefs on new evidence.

High NA → everything feels surprising, transition model is over-weighted. Seen in: PTSD, anxiety. Behavioral: hypervigilance, startle responses.

Low DA → flat policy posterior, indecision. Seen in: Parkinson's. Behavioral: bradykinesia, initiation deficits.

Low 5-HT → high policy precision on short-horizon preferences, impulsivity. Seen in: some forms of depression, OCD. Behavioral: inability to sustain long-horizon plans.

The book is careful: these are signatures, not diagnoses. The mapping isn't perfect. But every item is specific enough to wrong, empirical enough to check.

Why this matters for engineering

You don't need to be a neuroscientist to use this. The four precisions are parameters you set in any Active Inference agent.

observation_precision ≈ ACh. Workbench: adjusts weight on the likelihood term in Eq. 4.13.
transition_precision ≈ NA. Workbench: weight on B in the forward/backward messages.
policy_precision ≈ DA, aka γ. Workbench: preference_strength in the spec's Runtime block.
preference_precision ≈ 5-HT. Workbench: inverse-temperature of C.

When you tune these in the Builder or tweak preference_strength in a cookbook recipe, you're operating the same knobs the theory claims biology operates. Session 5.3 makes this explicit.

The runnable demo

/cookbook/predictive-coding-precision-gates-error is the recipe built around this session. Same agent architecture, two precision settings, two worlds. You watch:

Low observation-precision agent → ignores sensor, sticks to priors → hallucinates "I know where I am" even when the world shows otherwise.
Low policy-precision agent → can't commit to a single plan → hesitates, samples across multiple policies.
Low preference-precision agent → doesn't care enough about the goal → wanders.

Three pathologies. Three knob settings. The Workbench can simulate the behavioral signatures Session 5.3 describes.

The concepts this session surfaces

Precision — inverse variance.
Neuromodulator — a brainstem chemical that broadcasts a precision signal.
Signature — behavioral pattern predicted by a specific precision setting.
ACh/NA/DA/5-HT — the four modulators with the clearest mappings.

The quiz

Q: An Active Inference agent behaves as if it "can't commit to a plan" — its action posterior is diffuse even when the goal is clear. Which precision is most likely too low?

☐ Observation precision (ACh).

☐ Policy precision (DA, γ in Eq. 4.14). ✓

☐ Transition precision (NA).

☐ Preference precision (5-HT).

Why: The policy posterior is a softmax over −γG. With low γ, the softmax spreads across policies. The agent's behavioral signature — diffuse action selection with clear goals — matches low-DA clinical syndromes like Parkinson's. The other precisions affect different aspects of the loop (perception, dynamics, preference strength).

Run it yourself

/learn/session/5/s3_neuromodulation — session page.
/cookbook/predictive-coding-precision-gates-error — precision sweep.
/cookbook/planning-softmax-temperature — γ sweep ≈ DA sweep.
/cookbook/preference-precision-vs-strength — 5-HT ≈ preference-precision sweep.
/cookbook/perception-noisy-sensor-robustness — ACh ≈ observation-precision sweep.

The mental move

Chapter 5's boldest move is making clinical claims that depend only on the math, not on additional hand-crafted biology. If the theory is right, every computational-psychiatry finding should eventually map onto one of the four knobs above. That mapping is ongoing empirical work, not settled fact — but the framework is falsifiable and the predictions are specific.

Part 31: Session §5.4 — Brain map. Chapter 5's closer. A labeled diagram of the cortex with the generative-model factors drawn onto it. What areas do what. Where the theory's cleanest predictions live.

⭐ 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 29: Session 5.2 · Part 30: Session 5.3 (this post) · Part 31: Session 5.4 → coming soon