Active inference and different types of meditation
Introduction
I recently came across a nice paper that synthesizes, combines and improves existing active inference accounts of different types of meditation.
The paper, which is a master's thesis, is called Opening the Sensory Gates A Predictive Processing Account of Meditation and Learning. It was published in 2021 by Shawn Prest, a “PhD researcher working at the intersection of meditation, consciousness and computational neuroscience at the M3CS.”
The paper talks about other things too, but in this post I'm mainly going to present how it accounts for focused attention, open monitoring and insight meditation.
I'm not going to present here what active inference is (I'm lazy), but honestly, the paper provides the clearest exposition I've ever read. The section in which it summarizes active inference is less than 10 pages long, and I can only encourage you to go and read it, as it really does contain everything you need to understand the basics of the framework.
Focused attention
Roughly speaking, FA (focused attention) meditation consists of trying to keep our attention on something (often the breath); when our attention drifts (mind wandering, other sensations, etc.), we need to notice as quickly as possible that we are distracted, unstick our attention from the content of the distraction, then redirect our attention to the meditative object (the breath).
According to the paper, “non-action” and “letting go” are two important things related to FA meditation.
Non-action is about remaining non-reactive towards distraction. We must not engage with the distraction, but simply notice it calmly.
Letting go consists in gradually dropping more and more subtle forms of clinging. This is a bit hard to explain (it'll become clearer in the rest of the post), but the idea is that clinging is a kind of attachment to how one perceives experience, and this causes both a certain form of “mental tension” (negative valence) and the “fabricated” character of experience; eliminating these increasingly subtle mental tensions is a relief and makes perception less and less fabricated, with a kind of “fading” of perception.
In an active inference framework, the author argues that FA meditation does three things.
The first thing involves non-action. By practicing non-action, we refuse to engage with action policies arising from goal priors (in this case, distractions). This results in (1) greater control over the precision of goal priors and action policies (more volitional control), and (2) a reduction in active inference in favor of perceptual inference.
Having an HGM (hierarchical generative model) more oriented towards perceptual inference rather than active inference is useful for insight meditation in that it makes the model easier to update (more on this later in the post).
The second thing involves letting go. The idea here is to consider that letting go consists in reducing the precision of the upper layers of the HGM (as discussed in this post).
The fact that letting go leads to a fading of perception is explained by the fact that gradually dropping the upper layers leads to a less and less fabricated experience.
The fact that letting go is seen as the elimination of a tension is explained by the fact that, in active inference, prediction errors are negative valence; eliminating a layer reduces prediction errors in that there is no longer a conflict between the predictions of this layer and the data of the previous layer.
The fact that the tension is seen as being related to a form of attachment to how one perceives experience is explained by the fact that this is literally the case; the layers have their precious theories which they seek to impose in the system.
Non-action and letting go are mutually supportive. Non-action reduces active inference in favor of perceptual inference, which favors clearer mind, evidence accumulation and noticing mental tensions rather than being trapped in policies. Letting go increasingly deprives the system of the ability to generate policies by dropping more and more layers.
The third thing involves mindfulness. To be able to practice FA meditation effectively, you need to be able to realize whether you're distracted or not. So the idea is that, in FA meditation, we increase the precision of the meditative object, and in addition, we increase the precision of metacognitive inferences about whether we are focused or not.
Open monitoring
Roughly speaking, OM (open monitoring) meditation consists of non-judgmental observation of what appears in consciousness, letting it come and leave.
The paper seems to treat OM meditation with much the same “elements” as FA meditation, but in “different proportions”. One could say that FA meditation puts a lot of emphasis on mindfulness, and a little less on non-action, while OM meditation puts a lot of emphasis on non-action, and a little less on mindfulness. (Both probably put equal emphasis on letting go.)
In OM meditation, mindfulness is obviously still important to stay on track with the task, but this mindfulness is in a sense a little softer and less demanding, because we can let our attention wander from one sensation to another. The main point of the task is to observe in a non-judgmental way (non-action) the sensations that emerge.
In FA meditation, non-action is obviously still important to avoid getting trapped in policies, but non-judgmental observation is not the central point of the activity; noticing a distraction should indeed be done in a non-judgmental way, but we must not remain in our observation, we must redirect our attention to the meditative object.
Insight meditation
Roughly speaking, insight meditation is a form of deep introspection aimed at achieving a transformative shift of consciousness through a direct understanding of profound truths about experience, or by undoing deep delusions. In Buddhism, it can be the direct and transformative understanding of the three marks of existence.
The paper considers that practicing/succeeding in insight meditation requires a great deal of “concentration”, where concentration is understood as a combination of mindfulness, non-action and letting go.
With this concentration, attention is efficiently directed to a minimally fabricated experience accompanied by a top-down increase in the precision of the features to be noticed (such as the three marks of existence) to gather evidence for them.
Through this, the meditator will experience a “Bayesian model reduction”, which is the feeling of insight: “Essentially, Bayesian model reduction evaluates the evidence of reduced forms of a parent or full model by eliminating redundant parameters. Crucially, Bayesian model reduction can be applied to the posterior beliefs after the data have been assimilated. In other words, Bayesian model reduction is a post hoc optimization that refines current beliefs based on alternative models that may provide potentially simpler explanations. Previously, we have focused on optimizing free energy with respect to the (approximate) posterior as encoded by its expectations. However, we can also minimize free energy with respect to the priors, thereby eliminating redundant parameters to reduce model complexity.” (see)
Having a Bayesian model reduction may not be “totally transformative”, in that different contradictory models may remain in the HGM. Only when the new understandings come to dominate over the previous ones at a very high system-wide level can we speak of “perception-altering model reduction”.
Conclusion
All in all, the paper is interesting. I recommend reading it, it also touches on things I haven't mentioned, like opacity, neuroplasticity, autism, etc. I also plan to read the other papers (which sound fascinating) from this author and maybe make posts about them.
