lecture 2 - predictive processing

Question 1

predictive processing

Answer 1

• overarching theory that brains minimize prediction errors
• perception is a process of top-down inferencing instead of bottom-up inferencing
• active processing instead of passive processing
• example: when drinking cream soda instead of the expected coca cola, with active processing you get a PE and surprised response, with passive processing you just taste the cream soda without surprised reaction.

Question 2

what does the brain do according to predictive processing theory

Answer 2

everything the brain does can (in the long term) be explained as PE minimization

Question 3

PP explanation: binocular rivalry (house/face)

Answer 3

• BR occurs when the eyes are presented with different stimuli and subjective perception alternates between them
• PP explains this as:

1. the brain has priors for seeing a face or a house separately, but there is no prior for seeing both simultaneously in the same location, because objects can’t coincide in space and time
2. The predictive model can’t strongly favor one interpretation as no single model/hypothesis about the causes in the environment has both high likelihood and high prior probability.
3. As one stimulus (say, the face) becomes dominant, the brain “explains away” the perception of the face, but the other stimulus (the house) still exists as a suppressed prediction error (something that remains unexplained). The brain recognizes this error and may then switch to perceiving the house.
4. This switching continues back and forth because no single model fits the data entirely.

Question 4

stubbornness of priors

Answer 4

• priors are determined by past experience
• they are believed to be relatively hard-wired depending on the extent to which they are evolutionarily inherited, grounded in a lifetime of learning, or were acquired over a much shorter time scale.

Question 5

active inference

Answer 5

• the brain continuously predicts the outcome of its own actions to confirm/test its model of the world
• actions are driven by predictions and sensory PE minimization as well

Question 6

PP: sense of self

Answer 6

• in order to interact with the world effectively, our brain must have an internal model of ourselves as a cause in the world.
• This allows us to predict how our actions will affect the world around us and, in turn, how the world will affect us through sensory feedback.
• this renders our sense of self as an agent a construct or model of the brain. the “self” we experience isn’t necessarily an inherent or fixed entity but rather a mental construct.

Question 7

PP: the inside comes first

Answer 7

• the brain actively creates its own reality from the inside: it predicts and interprets the outside world continuously on the basis of its model
• influences from the outside are also critical in shaping model formation, but, the internal comes first

Question 8

top down influences on perception: PP vs traditional view

Answer 8

• traditional view: top-down processes in the form of memory, cognitive control, attention, etc. remain secondary in dealing with influences from outside.
• PP perspective: PP perspective, top-down influences are the primary influence on perception. external influences become secondary to anticipatory states.

Question 9

PP: tickling example

Answer 9

• you cannot tickle yourself as these sensations were predicted in advance. no prediction error = no tickling sensation
• schizophrenics can tickle themselves, as they have problems predicting the outcome of their own actions - making the distinction between internally and externally generated brain activity blurry

Question 10

PP: phantom pain example

Answer 10

• amputees often experience pain in the missing limb, despite its absence
• When a limb is lost, the brain still maintains this predictive model because it’s been conditioned by years of sensory data coming from that limb (stubbornness of priors)
• This supports the idea that our sense of self and bodily experience are not direct reflections of reality, but rather constructed by the brain through predictions. i.e., pain is a learned expectation or construct of the brain

Question 11

counterfactual predictions

Answer 11

• the deep temporal and hierarchical structure of generative models allow for consideration of the outcome of multiple actions, the further one goes into the future, without engaging in overt action
• active inference therefore includes ‘what if’ beliefs (counterfactual hypotheses) about the world, and belief updating that does not entail overt action (i.e., mental actions)

Question 12

learned cognition

Answer 12

• in the PP framework, perception, action, etc. are constructed through increasingly more reliably predictive models that reduce PEs
• therefore, past experience/learning is a pervasive factor underlying all mental activity
• our mental landscape is thus dominated by models that have reliably reduced uncertainty in the past

Question 13

mental action

Answer 13

• covert cognitive processes the brain performs to continuously update and refine its predictions.
• they are mental computations the brain uses to minimize uncertainty and prepare for more effective future actions.

Question 14

learned cognition: habitual vs goal directed mind

Answer 14

• learned cognition raises the question of how much of our thinking and behavior is automatic (habitual) versus goal-directed (deliberate)
• If most of our mental models are based on past experiences, much of what we do may be habitual, driven by well-established predictive patterns, rather than actively goal-directed.

Question 15

learned cognition: plasticity of the predictive mind

Answer 15

• concerns the brain’s ability to adapt and change its predictive models based on new experiences.
• how flexible or plastic is the brain in updating its models, and can we overcome habitual patterns to adopt new, more effective ones when necessary.

Question 16

plasticity and meditation

Answer 16

• meditation brings one into the present moment, which suggests a temporary suspension of predictive processing at certain brain levels
• meditation’s positive effects can be explained as weakened habitual perceptions and responses, so that attention is freed from conditioned patterns
• so, meditation gradually reduces counterfactual predictions/temporally deep cognition until all conceptual processing falls away, unveiling a state of pure awareness

Question 17

PP: confirmation bias

Answer 17

• PE minimization explains confirmation bias as looking for information that confirms our beliefs, which would then result in small PEs (model confirmation)

Question 18

PE: placebo/nocebo effects

Answer 18

• both phenomena are deeply tied to the brain’s expectations about what will happen as a result of the “treatment.”
• PE minimization explains this as the brain not just passively receiving sensory information but actively constructing perceptions based on its predictions

Question 19

PP: model of perception as controlled hallucination

Answer 19

• means that our brain is constantly generating predictions (or models) of what we expect to perceive, and these are updated or constrained by sensory input from the outside world
• if the brain’s predictions (the internal model) are not properly restrained or corrected by sensory input from the outside world, it can lead to faulty perceptions or behaviors such as OCD.

Question 20

summary of what the brain does

Answer 20

• continuously build models of the outside world on the basis of our interactions with the world to be able to predict what the world likely looks like now: to get a grip on the outside world, to reduce uncertainty, and to fit expected states
• in this perspective, the brain doesnt simply process information from the outside, but continuously generates information to meet internal expectations and to ensure the fitness of its internal models
• cognition is action-oriented ad probabilistic/predictive
• learning is a pervasive feature of all mental activity

Question 21

PP objections

Answer 21

1. dark room problem
2. mathematical formalization
3. not a neurophysiological theory
4. falsifiability

Question 22

PP objection: dark room problem

Answer 22

• If minimizing prediction error were the sole driving force of cognition and behavior, organisms should prefer to seek out highly predictable environments where there’s little to no change or uncertainty, like a dark room.
• answer: the brain is not a brain in a vat, but the brain (and hence the mind) is embodied within an organism that interacts with the environment. this suggests that organisms do not simply seek to minimize prediction error in isolation but do so in the context of their overall well-being, which includes interaction and engagement with the world.

Question 23

PP objection: mathematical formalization

Answer 23

• there are multiple versions of PP models, as well as various ways to formulate the free energy principle. there are mathematical inconsistencies in the formulations of these theories
• current versions of these models might be insufficient to explain the complex processes they seek to describe

Question 24

PP objection: not a neurophysiological theory

Answer 24

• the free energy principle lacks explanatory power as it is too abstract and doesn’t adequately connect with the actual biological and neurophysiological workings of the nervous system.
• it is a computational theory, not a neurophysiological theory (explains cognition and perception through mathematical and information-theoretic principles but doesn’t delve into the neurophysiological details of how neurons, brain circuits, and other biological structures perform these functions)

Question 25

PP objection: falsifiability

Answer 25

• The free energy principle and predictive processing aim to explain a broad range of phenomena—essentially all cognitive and behavioral processes—through a single unifying framework
• one of the issues with theories that claim to explain everything is that they are often difficult to falsify: it is hard to pinpoint specific predictions that could easily be proven false
• despite its breadth, the theory does generate concrete predictions that can be tested, which partially addresses the falsifiability issue.

Question 26

four neural predictions for PP (Walsh paper)

Answer 26

1. error-signalling neural responses to sensory stimuli should scale inversely with expectation
2. top-down signals represent sensory prediction
3. at each level of the cortical hierarchy there are two functionally distinct subpopulations representing predictions and prediction errors
4. prediction error minimization is achieved through repciprocal exchange of error and prediction signals across levels - a process known as ‘hierarchical inference’

Question 27

Walsh prediction 1: error-signalling neural responses to sensory stimuli should scale inversely with expectation

Answer 27

1. omitted responses: when expected sensory inputs are omitted, the brain still generates a response. this indicates that constantly generating predictions about the world, and its activity reflects this anticipation, even when the stimuli are absent.
2. repetition suppression: when the same stimulus is repeated multiple times, neural responses decrease. this is because the brain anticipates the repeated stimulus, reducing the need for a strong response (prediction matches the reality). If a new stimulus is introduced, the neural response increases again, reflecting the brain’s sensitivity to unexpected or novel input. however, repetition suppression could also be due to neural adaptation.
3. expectation suppression: when a stimulus is expected (and not omitted), the neural response is weaker because the brain is predicting it accurately, minimizing error signaling.
4. precision and attention: attention to a specfic stimulus can increase the brain’s prediction precision and the neural response. when you focus on a stimulus, the brain considers the incoming sensory data to be more important, upregulating the response

Question 28

challenge to Walsh prediction 1: error-signalling neural responses to sensory stimuli should scale inversely with expectation

Answer 28

• many of these effects have not been replicated due to different methodologies

1. global measures (fMRI, M/EEG) vs. direct neuronal recordings: the two neuronal populations and processing levels differ across these methods. expectation and error units are active at the same time and undergo distinct modulations. In global measures, these distinct signals may be averaged together, obscuring specific effects. Direct recordings may reveal clearer distinctions between these neural populations.
2. variation in the paradigms used: how predictions are instilled: length of training, probabilities, passive viewing/attended stimuli, implicit/explicit

Question 29

Walsh prediction 2: top-down signals represent sensory prediction

Answer 29

• neural activity believed to represent prediction appears to carry stimulus-specific information, which is heavily experience dependent, and interacts with bottom-up sensory input.
• predictive signals descend from higher-level processing areas

Question 30

challenges to Walsh prediction 2: top-down signals represent sensory prediction

Answer 30

1. while the theory suggests that top-down signals are predictions, it’s still unclear how the brain generates these predictions at a neural level. what exact mechanisms or processes extract patterns from past experiences to inform future predictions?
2. is the formation of sensory predictions a unitary neural process or an array of independent, task-tailored mechanisms

Question 31

Walsh prediction 3: at each level of the cortical hierarchy, there are two functionally distinct subpopulations representing predictions and prediction errors

Answer 31

• forward projections: PE signals (bottom-up)
• backward projections: predictions (top-down)
• forward and backward projections originate from separate cell populations
• only recently addressed empirically
• supported by direct neuronal recordings
• first fMRI studies are promising (i.e., deep layers contain expectation units, superficial layers contain error units), but not conclusive, as it cannot distinguish between different layers of the cortex

Question 32

Walsh prediction 4: PE minimization is achieved through reciprocal exchange of error and prediction signals across levels (hierarchical inference)

Answer 32

• hierarchical processing: there is evidence that the sensory brain is hierarchically organized, and that it exploits descending predictive activity to render sensory signals into meaningful constructs
• hierarchical inference: although the brain is clearly organized hierarchically, we have less empirical evidence showing how predictions and errors are exchanged and updated across these hierarchical levels in a way that predictive processing theory suggests

Question 33

conclusion of the walsh paper + future directions

Answer 33

• overall, empirical support is mixed for each of PPs key hypotheses, but there is no striking clear-cut counterevidence
• traditional theories of perceptual processing have failed to adequately explain why sensory cortices are infused with masses of descending connections. predictive processing provides such an explanation
• future studies should explicitly test hypotheses that are unique to PP, in particular provide definitive evidence of the existence of expectation and error units in neural processing

Question 34

main challenges of the Walsh paper

Answer 34

• limitations inherent to fMRI and M/EEG
• how to translate the algorithm into a neurophysiological model