Predictive Processing and Anxiety and other Maladies

This post is based on a paper written by Fabienne Picard and Karl Friston, entitled: “Predictions, perceptions, and a sense of self,” that appeared in Neurology® 2014;83:1112–1118. Karl Friston is one of the prime authors of predictive processing and Fabienne Picard is a doctor known for studying epilepsy. The ideas here are not new or even new to this blog, but the paper and specifically the figure below provide a good summary of the ideas of predictive processing. Andy Clark’s Surfing Uncertainty is the place to go if the subject interests you.

The brain is regarded as a predictive machine that continuously seeks to improve its beliefs or hypotheses through trial and error. In predictive processing, signals ascending from the sensory cortex (or subcortical structures such as the thalamus) are compared with descending predictions to form prediction errors. These prediction errors ascend to higher cortical regions to update posterior expectations and improve top-down predictions. In this scheme, the predictions are the stuff of perception, with the ascending prediction errors providing continuous feedback, which confirms or disconfirms the predictions or perceptual hypotheses. The figure taken from the paper by Picard and Friston provides a schematic overview of predictive coding in the brain. The neuronal message passing among different levels in cortical hierarchies comprises 2 streams: a descending stream of top-down predictions and an ascending stream of bottom-up prediction errors.


Picard and Friston state that the connectivity implicit in forming and accumulating prediction errors is consistent with known  micro-circuitry in the brain. Minimizing hierarchical prediction errors provides an internally consistent explanation for how sensations are generated—at multiple levels of abstraction or description. Heuristically, when the internal processes generating predictions in the brain match the external processes generating sensations, prediction error is minimized. By minimizing prediction error, internal processes therefore emulate or represent external processes. Statistically speaking, this corresponds to Bayesian inference about how sensory data are generated.

Accurate predictions suggest that cortical hierarchies embody a model of how sensations are produced or generated. This model is known as a generative model and has to be learned or acquired through experience. In other words, generative models have to be formed (via experience-dependent plasticity) to enable perceptual inference.  This acquisition—through the associative synaptic plasticity of recurrent connections in cortical hierarchies—underlies perceptual learning. Picard and Friston note that it is interesting that both perceptual inference and learning are driven by the same imperative—to minimize prediction errors.

In vision, an object can be inferred before it is completely disclosed, given a sufficient matching
between the prediction and early or partial visual cues. Should matching fail, creating a prediction error, an improved or updated prediction—that produces a smaller prediction error—will emerge. But how does the brain select among the myriad of prediction errors that it could explain away? In predictive coding, the precision of prediction errors reflects the degree of confidence in the information conveyed by ascending prediction errors at any level of the
hierarchy. Precision is the inverse of variability or uncertainty, so precise prediction errors at one hierarchical level will come to dominate inference in the remaining levels. A simple analogy here would be the confidence we place in—or weight we place on—information from a reliable or precise source. Thus the expected precision of different sources of (sensory) information can be associated with attending to a particular source. Psychologically, when one source (sensory modality or hierarchical level) is expected to have a higher precision than others, its precision increases and we attend to—or select—this source to update our predictions. Physiologically, this precision weighting is thought to be implemented neuronally by a top-down modulation of the gain of cells reporting prediction errors.

This means that ascending prediction errors are precision-weighted; in other words, a mismatch with a high expected precision will create an amplified prediction error (c.f., attentional gain). Put simply, ascending precision-weighted prediction errors report what is newsworthy in the incoming signal—not the sensory signal per se. The cortex is therefore more of a (salient) error detector than a feature detector. The notion of precision is an important aspect of predictive coding and suggests that the brain not only predicts the content of sensory input but also requires top-down predictions of the precision or confidence that contextualizes that input. It is the failure of this neuromodulatory contextualization that many people now consider a candidate target for such issues as epilepsy and anxiety.


In addition to perception and cognition, the predictive processing model also applies to action and offers a unifying model for perception and action, which have been regarded classically as distinct functional systems. In the predictive formulation, action is regarded as the fulfillment
of descending proprioceptive predictions through classical reflex arcs. In more detail, the brain
generates continuous proprioceptive predictions about the expected location of the limbs and eyes that are hierarchically consistent with the higher goal of the movement. In other words, we believe that we will execute a goal-directed movement and this belief is unpacked hierarchically to provide proprioceptive and exteroceptive predictions. These predictions are then fulfilled automatically by minimizing proprioceptive prediction errors at the level of the spinal cord and cranial nerve nuclei.


A large number of neurologic and psychiatric syndromes can be cast as a failure of neuromodulation or gain control.  Delusions and hallucinations can be regarded as false inference in the perceptual domain, while various psychomotor limitations can be regarded as a failure of descending motor predictions to elicit movement. All these instances of false inference can be explained in a simple way in terms of a failure to estimate the confidence in, or precision of, prediction errors. The nice thing about this is that it provides a functional explanation for the emergence of clinical symptoms while furnishing a neurophysiologically plausible explanation in terms of abnormal neuromodulation (possibly involving dopamine). According to Picard and Friston the most likely primary deficit is a failure of sensory attenuation. For example, during action we normally attenuate the sensory consequences of our own movement (for example, we are not aware of the optical flow produced by eye movements). Sensory attenuation causes us to underestimate forces we produce in relation to the same force that is applied passively—this is the force-matching illusion. That is, subjects overestimate the force applied by an external target when matching it. Effectively, this is because we pay less attention to self-caused sensations. Crucially, patients with schizophrenia show a failure of sensory attenuation and are resistant to the force-matching illusion. People with schizophrenia (and possibly autism) cannot ignore the sensations they cause themselves. It has been proposed that this failure drives compensatory increases in precision (postsynaptic gain) at higher levels of the hierarchy—at the price of explaining sensations in a delusionary fashion (i.e., by placing too much confidence in their prior beliefs). In short, a neuromodulatory deficit at low (sensory) levels of the hierarchy results in a failure to contextualize—or attend to—incoming sensory information. Compensatory changes in precision at higher levels of the cortical hierarchy endow hierarchical beliefs with too much confidence, thereby providing a Bayesian account of delusions and hallucinations.

The key point here is that apparently disparate clinical phenomena are underpinned by one simple mechanism—false inference due to aberrant encoding of precision through a neuromodulatory failure of synaptic gain control.

anxietyThe aberrant modulation of (attention to) prediction errors may also be associated with anxiety disorders. In line with this, comedienne Aparna Nancherla, described anxiety as having a particularly edgy improv group in one’s head that needs just one word to spin countless scenarios that no one is comfortable with. And none of those thoughts has a future. As she said, “I am too aware of my body,” and that might translate into giving too much precision to interoceptive cues while ignoring other cues. Having experienced anxiety disorder, the spiraling of slightly crazy ideas just requires precision to be a little off for then the improv group goes to work and to make each scenario a little crazier than the one before.

Picard and Friston have postulated that a pathologic absence of precise prediction errors may lead to feelings of trust, well-being, and inner peace, such as those encountered in some forms of epilepsy with “ecstatic” symptoms. This state could be understood as a disruption—by epileptic discharges in the insula —of the “comparator between predictions and real inputs” for multisensory integration and interoception. This could result in an “ultimate stable state” in which top-down signals would perfectly predict representations in lower cortical levels, which would be the state the brain is trying to achieve: a perfect prediction of the world, like the Oriental nirvana.