Predictive Processing
Active Inference
Embodied Cognition
Generative Mental Model
Predictive Coding
lots of fancy names, but it comes down to
the system works like this:
not passive reception
active construction
your perceptions ARE your predictions
until proven wrong
forward connections carry ONE thing
backward connections carry PREDICTIONS
this is where it gets interesting
error units - superficial pyramidal cells
encode precision-weighted prediction errors
representation units - deep pyramidal cells
encode predictions about causes
functionally distinct populations
same level, different jobs
the prediction travels DOWN
the error travels UP
when prediction matches input?
error units go quiet
explaining away
when prediction misses?
error signal propagates
demands attention
forces update
neural connections BACK 4 to 1
four times more backward than forward
WHY
neurons are EXPENSIVE
2% body weight
20% of your energy budget
why waste that on redundant bottom-up processing
evolutionary answer:
don't send what can be predicted
only send SURPRISE
only send ERROR
only send what matters
the brain implements efficient coding
bandwidth is precious
prediction errors are compressed information
this architecture answers:
predictive brain = metabolically efficient brain
all levels simultaneously
all the time
no central coordinator
not all errors matter equally
brain estimates reliability of prediction errors
high precision = high gain on error units
low precision = low gain on error units
neuromodulators encode precision
attention = precision optimization
you're not amplifying signal
you're changing confidence in prediction error
visual input is 2D
perception is 3D
that extra dimension?
generated by your model
photons hit retina
your percept = maximum a posteriori estimate
Think of visual illusions.
They aren't failures of the system.
They are optimised solutions based on
environmental stimuli and past experience.
throw something
catch something
you don't calculate trajectories
you don't compute joint angles
you don't plan muscle activations
you PREDICT proprioceptive consequences
then your body minimizes prediction error in the motor domain
prediction: arm here, ball arriving there
sensory feedback: arm NOT here yet
error signal triggers movement
movement continues until error resolves
perception: adjust predictions to match world
action: adjust world to match predictions
same computational principle
opposite direction of fit
like we are the puppet and the puppeteer
both at once
when you miss?
massive prediction error
attention increases
precision weights adjust
repeated action
learning = updating model based on precision-weighted error
This makes even more sense when you think about the system priority - Survival
imagine processing every photon
every sound wave
every tactile input
from scratch
every millisecond
metabolic bankruptcy
evolutionary solution:
build models
predict input
process only deviations
BANG!!!!
your system predicts: TIGER
massive error signal
attention spike
action cascade
turns out - Dave dropped something
system updates:
if you think you don't do this - bad news
you're descended from ancestors who predicted tigers in bushes and ran,
not from ancestors who carefully evaluated each stimulus and got eaten
false positives? survived
false negatives? eliminated
your threat detection is biased
by design
not by malfunction
Imagine you are walking along about to stand on a nail…
current model
sensory input arrives
prediction error generated
this is nociception
information about actual or potential tissue damage
NOT pain yet
the error signal is precise
increases gain on this error
changes model
compels action
subconscious level:
conscious level:
automatic - not a choice
new model assembled:
action follows:
perception/action loops
updated by prediction errors
middle state:
but also
new sensory information arrives:
new prediction errors generated:
precision weighting shifts:
model updates:
off we go
this is normal acute pain
error-driven updating
model plasticity intact
system functioning
what happens here
same injury
same healing process
same objective improvement
less certain information
OR
new information, less weighting
due to:
system maintains danger weighting
downgrades safety information
uncertainty persists
precision estimation fails:
prediction errors exist
but don't drive change
model stuck
danger predictions maintained
protective actions continue
perception/action INCREASES pain prediction
actions fit this catastrophic model
mechanisms:
anything positive gets downweighted:
self-fulfilling prophecy:
not conscious psychological choice
not willful catastrophizing
biological variation in system function:
influenced by everything everywhere all at once:
pain shouldn't persist when tissue heals
but it does
predictive processing explains this:
high-level predictions:
these generate lower-level predictions:
which generate prediction errors:
model becomes self-maintaining
interoceptive predictions also critical:
emotional feelings = interoceptive inference
anxiety, fear, distress = prediction errors in interoceptive domain
anterior insular cortex:
all reinforcing
all self-confirming
chronic pain shows precision dysfunction:
too much precision on:
too little precision on:
this isn't cognitive
this is built in system function
gain settings on error units:
neuromodulation potentially abnormal:
neural synchronization patterns:
precision weighting prevents learning
even when evidence available
signals don't have appropriate impact
if pain = maladaptive prediction
treatment = update the model
exposure-based approaches:
BUT - precision matters
if precision on threat too high
safety errors get ignored
exposure fails
need to:
context manipulation:
attention training:
pain education:
multisensory integration:
not "dealing with pain"
not "managing symptoms"
updating the generative model
motor control = active inference
you don't execute motor commands
you fulfill proprioceptive predictions
prediction: hand at cup
current state: hand not at cup
minimize error:
in motor domain - option 2
classical motor control:
active inference:
mathematically equivalent to optimal control
but mechanistically inverted
reflex arcs explained:
cerebellar function:
basal ganglia:
perception = inference about causes
action = inference through consequences
attention = precision optimization
learning = error-driven model updating
based on:
your experience of reality:
not what's out there
what your model predicts
constrained by sensory input
think of some patients
unhealthy, overweight, unemployed, smoking and drinking for Australia
chronic pain persisting
multiple failed treatments
every intervention somehow makes it worse
predictive processing lens:
their model predicts threat
their actions fit that model
context reinforces everything
influenced by everything everywhere all at once
precision weighting abnormal
model updating impaired
learning mechanisms disrupted
automatic
not conscious choice
the system doing what systems do:
minimize prediction error
given the priors it has
given the precision landscape
given the available evidence
bad things happen
models persist
suffering continues