Sandved-Smith, L., Hesp, C., Mattout, J., Friston, K., Lutz, A. & Ramstead, M. J. D. (2021). Neuroscience of Consciousness 2021(1):niab018. doi:10.1093/nc/niab018. Preprint: PsyArXiv, 2020, doi:10.31234/osf.io/5jh3c.
Key takeaways
From the reading, NOTE-958.
- Mental action as precision deployment. Attentional states “condition the precision of observations at the level below”; mental action selects transitions between them, “two possible actions u(2), ‘stay’ and ‘switch’”, at the second level only (Fig. 6 caption).
- Only precision, not content. The higher states “generate outcomes at lower levels that speak—not to which state the world is in but—to the precision of beliefs about which state the world is in” (Methods). The stimulus sequence is “predetermined, with a deviant stimulus shown every 20th time step” (Results, Level 1); precision acts as a “rate constant” on belief updating.
- Action one level inward. The true attentional state is a generative process with “a policy-dependent probability” of transition; the agent observes it (o(2)) and acts to observe itself “focused” (Results, Level 2).
- Meta-awareness and opacity. Meta-awareness is the precision of A(2); an attentional state is opaque when that precision is non-zero. Meta-awareness is set by hand in the simulation, high then low.
- Scope. No learning of any likelihood; single runs; no data fitted. The extension to other precisions (B, C, D, E, G) is stated, and the model “is not currently implemented to cover other kinds of mental action”.
Why this is here
The strongest form of CLAIM-337: active inference already has an internal loop, mental action, run as policy selection one level up. The reading confirms it more literally than the claim says. The paper models the agent’s attentional state as a generative process and has the agent act on it to fulfil a preferred self-observation, the action–perception cycle with the self in the place of the world. It narrows it too: one acting level, two actions, precision only, nothing learned. So it supplies none of the claim’s structure-learning strand.
For CLAIM-341 the registration asked whether the observation the mental action weights is fixed before the action. At the perceptual level it is: the stimuli are predetermined and mental action changes only their precision. At the attentional level it is not: the mental action changes the attentional state and so the content of the self-observation that follows, which is the synchronic form the thesis does not claim. The paper has no mechanism by which a judgement reshapes the observer’s likelihood.