In this post, Sophie Stammers reports from the Chandaria
Lectures, hosted by the School of Advanced Study at the University of
London. Professor Andy Clark, of the
University of Edinburgh, gave the annual lecture, where he introduced the notion
of ‘predictive processing’. Over the course of the three lectures, he
put forward the case for understanding many of the core
information processing strategies that underlie perception, thought and action as
integrated through the predictive processing framework.
On a model of perception
popular with Cartesians, and undoubtedly dominant in areas of the cannon that I
was acquainted with as an undergraduate, perception is something of a passive
business. Perceivers employ malleable receptor systems that (aim to) faithfully
imprint the world as it is, delivering a raw stream of information that is made
sense of downstream in later processing. Clark dubs this the “cognitive couch
potato view”. Despite its past popularity, this view seems incompatible with
evidence from multiple research streams in cognitive science which
indicate that perceivers are far from passive, and bring many of their own
expectations to the table. Predictive processing (PP) aims to provide a story
which both accounts for and unifies these findings, whilst also doing justice
to the human experience in the midst of it all.
PP systems don’t just take in
sensory information from the world, they are constantly trying to actively predict the present sensory
signals with use of probabilistic models. Incoming sensory signals are met by a flow of
top-down prediction, and when this matches the sensory barrage, the system has
unearthed the most likely set of causes that would give rise to the particular
experience. “Prediction errors” (information about mismatches between current
prediction and sensory information) indicate a gap in the predictive model, and
that a new hypothesis should be selected to accommodate the current sensory
signal.
Maybe, rich world-revealing perceptions – as of tables, chairs,
conversations, lovers, etc – only arise from the otherwise indiscriminate
sensory barrage when the incoming sensory signal can be matched with top down
predictions.
The PP system also weights
incoming sensory data. Top down prediction can dampen some of the sensory
signals whilst amplifying others. The “mask illusion” illustrates this nicely.
The system forefronts
“this-is-a-face” signals whilst treating signals that contradict that as noise. Those unfamiliar with this illusion
can watch celebrities grapple with their own predictive system weighting here. The illusion also seems to work with reptilian faces.
PP systems
are consistent with, but don’t necessitate “inbuilt” hypotheses – Clark assured
us that they may operate anywhere along the nativist-empiricist spectrum. Whilst
coming ready made with some innate hypotheses might give a predictive system a
head-start, we can get still get the story going with a system that generates
hypotheses over and over until a match with incoming sensory information is
made. Repeat the process for some other aspect of the world, until you hit on a
match…and so on.
Results from a number of fMRI
studies and research on attention are illuminated by the PP framework (e.g. Murray
et al., 2002; Muckli,
2010; Adams et al., 2013;
Schröger et al., 2015). For instance, when perceivers expect a
familiar sequence of birdsong to continue in a particular way, and when the
sequence is unexpectedly cut short, we see cortical activity consistent with a
momentary hallucination of the missing chirp, followed by a big burst of
prediction error (Adams et al., 2013).
On Clark’s account,
action flows from prediction as much as perception does. A simple motor action
is a matter of predicting the trajectory of proprioceptive sensations that
would ensue if you were to perform that action. This generates a flow of
prediction error (because you are not currently performing the predicted
action). The prediction error is resolved by down-regulating the impact of
sensory information specifying current bodily position, and performing
the action in question.
In many
cases, perception is directed at predicting how actions will unfold (Tatler et al. 2011). For instance, people gaze just
ahead of the knife’s point of contact with the bread as they cut a sandwich (Hayhoe et
al., 2003). Further, PP can account for “utility based skewing” in
which systems favour perceptions that best serve their actions (Mark et al, 2010). Clark
proposed that PP systems might be considerably geared towards action
generation, and that we should think of prediction error signals as sensory information that
hasn’t yet been leveraged to inform a rolling sequence of action engagement in
the world.
PP might also
explain something of the nature of our percepts. Clark drew our attention to a
range of evidence in support of the claim that prediction of a stimulus causes
it to enter conscious experience more readily, at which point it is dealt with
more efficiently than unexpected stimuli. Previous expectations of visual
stimuli determine when the stimuli in question enters conscious experience (Melloni et al. 2011). We also predict the intensity of pain on the
basis of prior expectations, which modulate the experience accordingly (Brown et al, 2008). Further, interoceptive prediction – that
based on sensing our own visceral states – plays a role in how we perceive
ourselves and others - estimations of our own cardiac states interact with our
perception of the intensity of another’s emotion (Gray et al. 2007). PP also fits with evidence that we bring
social and cultural expectations to our percepts, such as, for instance, when perceiving
an indiscriminate object as a weapon when in the hands of a person who fits our
expected racial profile of a gun wielder (as in Payne,
2006).
What about the big human stuff? Our goals, intentions,
motivations? Our grand projects? Clark floated the idea that it might be
“prediction all the way up”. When
you have some idea about how you might act in the world, then you can infer
information about how your sensory barrage would change in response to
performing those actions, and those predictions are what get you started. We
considered a big project, and how that might be broken down into sequences of sensory
matching and prediction error-driven action. For example, imagine that you want
to sail a yacht around the world. This leads you to predict you’re going to
pass your yacht master exam. That prediction itself entrenches how information
is processed at a more local level. You predict, for instance, that you’re
going to engage in flashcard revision when you can, and when you recognise an
appropriate opportunity, sensory modes engage flashcard practice. And so we go
on, installing expectations about our own movements though the world, and these
expectations themselves lead us to match incoming sensory information or
resolve prediction errors in accordance with those expectations.
PP is a big
theory promising big results. We were left with a flavour of what sorts of
questions future research is going tackle. What does all of this have to say
about the ‘hard’ question of consciousness, or, alternatively, might it
illuminate why we’re asking the wrong question there? What can PP tell us about
abnormal cognitive functioning? Does the current model combine top down and
bottom up processing in the right way, and how might this issue be tested? How
do social and cultural influences sculpt the brain, and can we use PP to do
better in these cases?
For those
who predict that they will be keeping up with future research, Andy Clark is
heading up a project entitled “Expecting
Ourselves: Embodied Prediction and the Construction of Conscious Experience”
(XSPECT) that will be investigating these and related issues over the next four
years. You can minimise
prediction error by visiting the project site
and staying up to date on the Brains
Blog.