Basically a high-level overview of the main contrast between the traditional cognitive science depiction of perception vs his theory.

  • Traditional: perception is a cumulative bottom-up process. you detect color and frequency, then edges, then curves, then combine that with stored knowledge and you perceive a coffee cup.

  • Predictive Processing: perception is prediction. Incoming sensory data is meshed with prediction, error signals propagate. When the flow of prediction adequately accounts for the incoming sensory signals, the coffee cup is perceived.

Other notables from the chapter:

  • He discusses the difficult task facing the brain; it sits in a dark box and all it has is a set of noisy sensors and it needs to make meaning from the signals of those sensors. Bootstrapping seems tough!

  • One of the benefits of the predictive processing paradigm: the situated state of cognition continually provides training data (you guess what’s going to happen next, then the universe reveals the answer).

  • Neural matter is incredibly dense. An MNIST network can fit into .002 cubic mm of mouse cortex.

  • The (comprehensible) universe is structured and compositional. Hierarchy in the brains is used to perceive this structure.

  • Hierarchical predictive coding: only the error signal (“surprisal”) at each level is propagated up to a higher level. This saves neural bandwidth.

  • Perception actually occurs when our predictions align with our sensory data.

  • Evidence is presented for “dynamic predictive coding” in retinal cells; the cells act to strip predictable elements from the visual stream so only the most noteworthy elements of the stream get forwarded up the hierarchy.

  • Evidence is presented for bi-stable binocular rivalry. If you show a subject two different images, one in each eye, they will generally alternate between perceiving one and the other. When you perceive image 1, error is minimized for that image but high error is coming from image 2, so you change your prior and minimize the error for image 2 but this exacerbates image 1 error. “Hyperpriors” are mentioned, priors that are deeply rooted/baked-in like the 2 different images can’t both be occupying your visual field.

  • Empirical bayes: your has priors, those priors get updated as evidence comes in to minimize error. Perception is minimization of sensory prediction error.

  • Predictive processing conceptualizes the forward flow as conveying error and the backward flow as conveying prediction that can “explain away” error. Thus it can both magnify or inhibit selectively.

  • Evidence for the bayesian brain is presented. A number of optical illusions can be recreated by using optimal Bayesian estimators. Hollow mask illusion occurs in neurotypical people but less so in schizophrenics (perhaps suggesting schizophrenia is related to broken priors or prediction engines.)

  • Predictive processing sees the brain as an active, generative thing (“proactive predictavores”) rather than a passive recorder.

THOUGHTS: Good to read at a high-level and then fill in the picture with targetted reviews. A bit confused, are there both representational and error neurons or is that all combined in the communication between layers. Brains are proactive predictavores.