Debate at a glance
- Summary: Whether shifting into a receptive, sensory, prior-updating mode (perceptual inference) constitutes a distinct regulatory strategy in its own right, or whether the very act of allowing/choosing that shift is itself an instance of active inference — a meta-level action on one’s own attention/precision-weighting — such that ‘regulation’ may always trace back to active inference, with perceptual inference only describing what unfolds once that shift has occurred.
- Status: open
Curator's take
Most people don’t really treat perceptual inference as a regulatory option at all — regulation tends to get equated with active inference by default. I’m not sure if that’s a real gap in how the field reads predictive coding, or if I’m the one misreading the model by suggesting there’s regulation through updating rather than only through action: is the choice to ‘let go’ — to shift into a receptive, sensory, updating mode — itself an active inference? If so, does that collapse the distinction entirely, or does it just mean active inference initiates the shift while perceptual inference describes what happens once you’re in it? This is a central question for me: the ability to move into that sensory, receptive, updating mode of perceptual inference — and what it even means for that move to be ‘chosen’ — is something I keep coming back to.
Positions in the field
- Norman A. S. Farb et al. (2015): Perceptual inference is presented as a genuine, underspecified second regulatory route alongside active inference — mapped onto contemplative acceptance/equanimity — but the paper itself flags that classifying it as ‘bottom-up’ rather than a cognitive/regulatory act ‘may be controversial,’ leaving the question explicitly open rather than resolved. (Farb et al. (2015))
Connections
- Related debates: inferential?, contemplative training increase objective interoceptive accuracy?
- Related concepts: Perceptual inference (interoceptive), Active inference, Predictive coding, Presence and agency
Raised directly in discussion of Farb et al. (2015)‘s active/Perceptual inference (interoceptive) distinction (see Norman A. S. Farb). The paper proposes perceptual inference — updating priors to match sensation, mapped onto contemplative acceptance and equanimity — as a regulatory route standing alongside Active inference. But a sharper question sits underneath that framing: is perceptual inference a regulatory strategy at all, in any sense that doesn’t ultimately bottom out in active inference?
The puzzle
If “regulation” means intentionally selecting among options to reduce prediction error, then even the decision to stop regulating in the active sense — to relax control, redirect attention toward sensation, and allow priors to update — looks like it requires an agent to do something: allocate attention, lower the precision it assigns to its own priors, permit an update. That “something” has the shape of an action. On this reading, perceptual inference never happens as a genuinely passive, unchosen event; it is always downstream of a prior active-inference-like act (the choice to let go), which would mean the two modes aren’t peers but are nested — active inference at the meta level, perceptual inference only describing the content of what follows.
Why the paper leaves this open
Farb et al. explicitly flag their own classification of acceptance as “bottom-up perceptual rather than top-down cognitive” as potentially controversial (see the notes field on Perceptual inference (interoceptive)) — but the paper does not resolve why it might be controversial. This debate names that gap directly: the controversy plausibly lives exactly at the meta-level question above, not merely in loose terminology.
A relevant absence in Seth & Friston (2016)
Seth & Friston (2016), the formal successor to Seth’s original framework, splits active inference itself into epistemic (information-seeking action) and instrumental (regulatory/control action) modes — but proposes no third, passive/perceptual mode alongside them. That silence cuts both ways: it could mean the authors simply didn’t address regulation-via-updating (an omission, not a rejection), or it could mean the architecture, worked through carefully by its own originators, has no natural place for perceptual inference as a peer category — lending some weight to the collapse-into-active-inference reading above. Recorded here as suggestive, not decisive, since the paper never engages the question directly.
Stakes
This bears on the wiki’s broader inferential? debate (is interoceptive processing ever genuinely bottom-up, or is everything, at some level, top-down prediction?) and on contemplative training increase objective interoceptive accuracy? (if “letting go” is itself a trained active-inference skill, that reframes what contemplative training is actually training — not raw sensory accuracy, but the meta-level capacity to initiate and sustain a perceptual-inference episode). It also touches Presence and agency: if presence (via perceptual inference) requires an antecedent act of agency to initiate, the two constructs may be less independent than Farb et al.’s “two complementary senses” framing suggests.
Status: open, and — per the my-take above — the live center of the author’s own current thinking on this material.