Book Review: Surfing Uncertainty

  • Predictive Processing is the unifying framework theory of how the brain works
  • Begins by asking, “How does raw sense data get turned into a unified picture of the world?”
    • Brain is a multi-layer prediction machine
    • 2 streams
      • Bottom-up stream of sense-data that is progressively filtered and abstracted
      • Top-down stream of predictions - starts with abstract models and generates predictions for lower layers
    • Both streams are probabilistic
    • 2 streams interact with each other at each layer of cognitive processing (using Bayes Theorem) to integrate the two sources of probabilisitic evidence
    • Possible outcomes of this interaction
      • Prediction and sense data match: models are good and layer stays quiet
      • Low-precision sense-data mismatches high-precisiion predictions
        • Prediction is probably more correct than sense data
        • Reinterpret sense data to match prediction
      • Unresolvable conflict between sense data and prediction
        • Signals “suprisal”
        • Alert that higher levels have to update their model
        • Surprisal signals for the stream of “low-level” data for higher levels
    • Perception is the result of these interactions
    • Top-down models explain why our brains can resolve pictures from noisy images, read words that have the letters jumbled, and automatically skip over grammatical errors
  • How does Predictive Processing explain pretty much everything about our perception?
    • Attention
      • Predictive processing measures the “confidence interval of predictions”
      • High-attention: perception is driven by bottom-up stream of sense data
      • Low attention: perception is driven by top-down stream of predictions from models
    • Imagination, dreaming, etc
      • The highest levels of our brain make predictions
      • As these predictions filter down into lower levels, they’re turned into specific sensations that we ought to be perceiving (which are then compared to incoming sense data)
      • Dreams and hallucinations are predictions that are either unconstrained or are weakly constrained by sense-data
    • Priming
      • Priming constrains the set of predictions, and thus makes it more likely that the brain will decode sensory data in a particular way
    • Learning
      • The brain is constantly creating models and generating sense-data predictions from them
      • Models that perform well are retained, while models that do not perform well are discarded
      • Hyperpriors - priors that one must have to form coherent models in the first place
        • Synchronicity of senses - different senses describe the same world
        • Object permanence - things don’t go away when we stop looking at them
      • While hyperpriors may be innate, infant behavior indictates that even hyperpriors may be learned
    • Motor behavior
      • Predictive processing isn’t something that informs our motor control, it is our motor control
      • Sometimes it’s easier to update the world to fit the model rather than updating the model to fit the world
      • The driver for us to move our bodies is a prediction that our bodies will be in a particular location
    • Placebo effect
      • Placebos operate much like priming described above
      • Bias the brain’s models to interpret sense data in particular ways
    • Neurochemistry
      • Predictive processing provides a coherent framework for what various neurotransmitters actually do
      • NDMA-glutamatergic system - top-down predictions
      • AMPA-glutamatergic system - bottom-up sense data
      • Dopamine - confidence intervals and other metadata
    • Autism & Schizophrenia
      • Both autism and schizophrenia can be interpreted as problems with updating models
      • Autism: models are too precise
        • Reality constantly deviates from models, generating surprisal signals
        • Explains why autistic people get irritated by minor things, like tags on clothing
        • Also explains why autistic people do well in certain fields that require high-precision models
      • Schizophrenia: models are too broad & agree with sense data too readily
        • More likely to assign significance to weak or noisy sense data
        • Aren’t fooled by optical illusions (which take advantage of models’ ability to overwrite sensory perceptions)
        • Schizophrenic delusions are the result of the brain assigning significance to every bit of sense data, no matter how minor or noisy
  • Predictive processing puts the rationalist project on a sounder scientific footing - shows the mechanisms by which our brains update; we just need to figure out how to make this system update better and faster

Predictive Processing and Perceptual Control

  • Predictive Processing is presaged by Powers’ Perceptual Control Theory
  • PCT is a good theory but it’s better seen as an approximation of Predictive Processing
  • If Predictive Processing does turn out to be the grand unified theory of the mind, PCT will be seen as an important precursor

How Do We Get Breasts Out Of Bayes Theorem

  • The one thing that Predictive Processing has trouble with is instincts
  • If the brain is one big prediction/verification machine, where do “hard-wired” instincts fit into that?
  • Maybe sometimes there’s enough of an evolutionary advantage to “hard-wiring” certain features (like recognizing beaks with red dots, in the case of seagulls) that it happens
  • Maybe instincts are the “default weights” assigned to models, and that most people go along with the defaults