Perceptual awareness and active inference

Study at a glance

• Authors: Thomas Parr, Andrew W. Corcoran, Karl J. Friston, Jakob Hohwy
• Year: 2019
• Journal: Neuroscience of Consciousness
• DOI: 10.1093/nc/niz012
• N: N/A (simulation / theoretical)
• Design: Computational modelling — active-inference (partially observed Markov decision process) simulations of two perceptual phenomena

Key findings

• Perceptual awareness depends on how an agent actively engages its sensorium; both overt action (saccades) and covert action (attention) ‘forage’ for the data that best resolve uncertainty, under the same variational free-energy machinery.
• Troxler fading — a peripheral percept dissolving during sustained fixation — is reproduced as the accumulation of uncertainty about a hypothesised stimulus once saccadic exploration is disrupted; because the generative model holds a prior that the world is volatile (imprecise state transitions), beliefs about unsampled stimuli decay toward maximum entropy.
• Binocular rivalry (and Necker-cube multistability) emerges from the identical mechanism with a covert attentional action rather than an overt movement: the attended feature is the one whose absence carried the most resolvable uncertainty last time-step, so attention alternates and the percept alternates with it — no eye movement required.
• Attention is modelled as (covert) action / gain control — a policy that sets the precision of a likelihood mapping — making ‘attention as action’ and ‘attention as precision’ the same move; alternation is driven by minimising expected free energy (epistemic value), not by stimulus fatigue.
• Individual differences in two precisions — the precision of transitions (how volatile the world is believed to be) and of likelihoods — reorganise the whole phenomenology, giving ‘distinct computational phenotypes that underwrite individual differences in perceptual experience’ and linking rivalry/fading susceptibility to aberrant-precision accounts of autism and schizophrenia.
• Yields testable predictions: time-to-fade should be inversely related to an individual’s estimated volatility (potentially readable from pupillary/noradrenergic dynamics), and the belief-updating at a perceptual switch should be measurable with EEG.

Limitations

• Purely theoretical / simulation: the authors state a comprehensive evaluation of the empirical data is beyond the paper’s scope, so the predictions are advanced but not tested here.
• Exteroceptive throughout (vision) — its relevance to this wiki is structural (the active-inference apparatus, and attention-as-covert-action), not an interoceptive result.
• Formulated in discrete time, which the authors note makes it hard to state definite predictions for the continuous dominance-time distributions the rivalry literature actually measures.
• The autism/schizophrenia predictions rest on aberrant-precision assumptions and, as the authors concede, could be confounded by independent attentional deficits in those conditions.

Connections

• Concepts: Active inference, Predictive coding, Interoceptive precision, Perceptual inference (interoceptive)
• Debates: inferential?
• Related studies: Friston (2005), Seth & Friston (2016), Seth (2013), Harrison et al. (2021), Petzschner et al. (2021)

Not an interoception study. It is kept structurally, the way Friston (2005) and Christoff et al. (2018) are: the wiki runs on Active inference everywhere, almost always meeting it in its interoceptive application (Seth, Barrett, Stephan, Petzschner). This is the machinery worked undisplaced, in vision, where two facts of the interoceptive account become visible in isolation — that attention is covert action, and that precision is what individual differences are made of.

The claim

Perceptual awareness “depends upon the way in which we engage with our sensorium.” Under Active inference, perception and action are one inferential process: actions are experiments that gather evidence for or against hypotheses about the causes of sensation. The paper’s move is to take two phenomena usually explained by passive mechanisms — a percept fading, a percept flipping — and show they fall out of what happens when an agent that must actively sample the world is prevented from doing so, or does so only covertly.

Two features of the generative model do all the work:

1. Action forages for disambiguating data. Whether skeletomotor (a saccade) or mental (an attentional shift), action is selected to minimise expected free energy, whose epistemic term rewards resolving uncertainty. This is the same epistemic/instrumental split Seth & Friston (2016) draw for interoception.
2. A prior that the world is volatile. State transitions are given finite precision (an inverse-temperature term ω), so in the absence of new data, uncertainty about a stimulus accumulates over time — beliefs relax toward maximum entropy. This is the “distrusting the present” idea (Hohwy et al. 2016) made mechanical.

Troxler fading as accumulated uncertainty

Fixate the central cross of Fig. 1 and the peripheral coloured blobs fade. In the model, holding fixation means the peripheral locations are not sampled; because transitions are imprecise, the posterior over each unsampled stimulus drifts to uniform — the percept “fades” because the belief about it has lost all its information. Resuming saccades reinstates the percepts. The paper reads this as compliance with Jaynes’s maximum-entropy principle: with no data feeding the ‘energy’ term, inference defaults to the maximally uncertain distribution. Crucially the fade is not retinal fatigue — it follows from the generative model’s volatility prior, which is why the fade-time should track an individual’s ω (Fig. 5), a quantity potentially estimable from pupillary/noradrenergic dynamics (Vincent et al.). That last link is the paper’s cleanest bridge back to this wiki: precision as a physiologically instantiated, individually varying gain — the same lever Interoceptive precision turns on.

Binocular rivalry as covert attentional action

Rivalry poses the sharper problem: the percept alternates with no overt movement. The paper’s answer is that attention is itself an action — a covert policy that sets which likelihood mapping is treated as precise (“how uncertain am I about what I would see if I attended to that?”). Directing attention to one eye’s stimulus lets uncertainty about the other accumulate, so the unattended hypothesis becomes the more epistemically valuable thing to sample next, and attention (hence the percept) flips. Rivalry is thereby “an attentional phenomenon” (Leopold & Logothetis; Zhang et al.) given a generative-model derivation. The same architecture, relabelled, produces Necker-cube reversals — “a common architecture for the neuronal message passing, even if implemented in different neuroanatomical structures.”

Why it earns a place on an interoception wiki

• Attention = covert action = precision. The wiki asserts this identity in several places — Predictive coding (“precision-weighting… interpreted as attention”), Farb (2013) (attention decides which cortex represents the breath), Interoceptive precision (precision is a dynamically allocated weight, not a trait). Parr et al. simulate it: attention as a policy selected to minimise expected free energy is precision-setting cast as action. It is the formal statement of the thing the interoceptive-attention studies show empirically.
• Computational phenotypes. The paper’s payoff is that fixed sensory input plus different precisions yields dramatically different perceptual experience, and it ties this to autism (Lawson, Palmer, Pellicano & Burr; van de Cruys) and schizophrenia — the same aberrant-precision lineage the wiki’s computational psychosomatics and Interoceptive precision pages invoke for interoceptive symptoms. It is the exteroceptive twin of the clinical story Harrison et al. (2021) tests in the body.
• Self-evidencing and the two authors it adds. The paper is built on Hohwy’s ‘self-evidencing brain’ and Parr’s active-inference MDP formalism, giving the wiki first-hand pages for two theorists it had been leaning on through Friston.

Held loosely

Everything here is simulation. The empirical hooks (fade-time ∝ 1/volatility; EEG at switches; pupillometry for precision) are proposals. And it is vision, not viscera — the interoceptive reading is a generalisation the wiki’s own sources (Seth, Quadt et al.) license but this paper does not itself make.