Week 6 Seminar: Prediction

The week the course turns over. Perception as mostly construction, and illusions as the mechanism working rather than failing.

Part II — Interior models1 April 2027Week 6 lectureSlidesSchedule

Everything in the course so far has flowed inward: world to sense organ to representation. This week it reverses. On predictive accounts the system generates what it expects and passes only the mismatch upward, which means most of what you experience was manufactured internally and checked against input rather than delivered by it.

Clark’s 2013 target article is the accessible statement and the reading for this week. Friston’s free energy formulation, from 2010, is the ambitious version.

What it explains, and what that costs you#

The reason this framing has taken over is that illusions stop being a list of curiosities and become predictions. You continue to see the Adelson checker-shadow wrongly after you have been told exactly how it works, and knowing does not help, because the inference happens below the level where knowing operates. Simons and Chabris’s gorilla is on the same footing.

Which leads to the sentence that makes this week the one people find hardest: you do not have access to which parts of your experience were sensed and which were generated. There’s no tag. Both arrive as perception.

The objection you are required to have#

Regarding the free energy principle, you are expected to arrive with a criticism rather than a summary. The strongest available is the dark room problem — if a system minimises prediction error, the optimal strategy is a featureless room and no movement, and the standard reply invoking expected free energy and novelty is contested. Litwin and Miłkowski’s “unification by fiat” is the sharper methodological version.

Grade the framework [C], and be ready to say what would move it. A theory that explains everything available is in a slightly awkward position, and saying why is part of this week’s work.

One distinction to get right#

Predictive coding is a specific proposal about neural implementation. Predictive processing is a broad framework. The free energy principle is a mathematical claim of much wider scope. These three get conflated constantly, including in published work, and keeping them apart is most of what it takes to read this literature carefully.

Reading#

The paper this seminar is built around:

• Clark (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences.

The rest of the week’s reading:

• Friston (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience — the ambitious version.
• Litwin and Miłkowski (2020). Unification by fiat: arrested development of predictive processing. Cognitive Science — bring this one to the seminar.
• Simons and Chabris (1999). Gorillas in our midst. Perception.
• Seth (2021). Being You. Faber — the accessible book-length treatment.

Grades are not attached to papers here, because a grade belongs to a claim rather than to the paper it appears in — see evidence grades. Several of these papers contain claims at two different grades.