Perception

Question 1

What does the predictive coding model of perception state?

Answer 1

The brain continually generates models of the world based on context and information from memory to predict sensory input

• top-down predictions against bottom-up evidence along the visual cortical hierarchy
• minimizes prediction errors
• we don’t have direct access to the world; we must use experience to infer

Question 2

What is bottom-up processing?

Answer 2

Information coming in from the world goes through multiple levels of processing to “make sense of it” by integrating it with prior knowledge

• perception based on sensation, not conceptual ideas
• mostly feedforward processes
• actual signals don’t get passed forward, only prediction error does

Question 3

What is top-down processing?

Answer 3

The use of context and general knowledge to understand and interpret sensory perceptions

• perceptions begin more generally and move towards specificity
• prior knowledge fills in the blanks to anticipate what is next

Question 4

What are priors?

Answer 4

Hypotheses about how likely things are in general and how likely they are to be true in the current situation

Question 5

What is the problem of perception?

Answer 5

• We only have direct access to the effect of the world on our senses, not to the world itself
• Cause (world) and effect (sensory interpretation) is not a one-to-one relationship; ie, one cause can produce different effects, and vice versa

Question 6

What is feedforward processing? Name a feedforward feature based model.

Answer 6

Sensory information encoded in early sensory areas is relayed from one node to the next; populations of neurons at each level respond to features of objects at an increasingly large scale and higher levels of abstraction

ie, information comes in, moving from basic sensory areas to high level integration areas of the cortex

• as opposed to feedback
• eg, visual processing is feedforward

Question 7

How does the brain decide which hypothesis/interpretation to apply to patterns of lower level features that are detected?

Answer 7

Integration of top-down information; ie, information generated by the brain to apply to the world

Question 8

What does a predictive coding model do? (3 steps)

Answer 8

1. The brain creates an internal generative model of the world with a prediction of what will be observed next
2. Expectations are projected down (eg from PFCs to visual cortices)
3. Input from the world is the error between expectation and received information (was the prediction right or wrong?)

Question 9

How do predictive coding models view the role of visual cortex neurons?

Answer 9

• predictions are based on the probability that the stimulus will have particular features
• error detection responds to a mismatch between predicted signal and actual signal

Question 10

What is the brain’s task in perception, according to the predictive coding model?

Answer 10

To predict the hidden cause out in the world of what we are perceiving

Question 11

What are the 4 steps (levels) of the predicting coding model?

Answer 11

1. Generative model: uses what we know about how the world works to generate predictions about what the object or scene in question is (eg where animals are found to estimate probability that there is one in front of you, given your location)
2. Higher level hypotheses (eg there are wild animals in the mountains)
3. Lower level hypotheses (eg if there is an animal, I should see movement and hear rustling)
4. Lowest level hypotheses: specific to each modality (touch, taste, smell, hearing, vision); hypotheses are compared with information input from the senses (eg if it’s an animal, eyes should be detected in the parts of visual cortex that pick up their contrast and shape; motion should be detected in MT)

Question 12

What are representational units in predictive coding?

Answer 12

Units at each level of the predictive coding model that encode expectation (the probability of a given stimulus under the circumstances, ie conditional probability); these units send predictions to the next lower level

Question 13

What are error units in predictive coding?

Answer 13

Units at each level of the predictive coding model that encode or read surprise (the mismatch between predictions and bottom up sensory evidence); sent forward to the next higher level, where expectations are adjusted or sent up to the next level

Question 14

How is prediction error generated and what does it do?

Answer 14

If error units send signals forward to the next higher level and there is a mismatch, prediction error is generated

The prediction error moves up the hierarchy, causing revision of hypotheses at the level above; if that level can’t minimize the prediction error, it is pushed up to the next level

Higher level = more substantial revision

When minimized, the winning hypotheses forms the contents of perception

Question 15

What did the Egner & Summerfield paper test?

Answer 15

Predictions stemming from predictive coding models against feature based models of object perception in the ventral stram

• used an encoding approach to fMRI

Question 16

What was the big picture question (Egner &Summerfield)?

Answer 16

Does predictive coding explain visual object recognition better than classic hierarchical feature-based models?

Question 17

What brain mapping knowledge did they leverage to answer a more focused question (Egner & Summerfield)?

Answer 17

• we know that category selective populations in the FFA respond more to faces than houses, and in the parahippocampal place area (PPA) more to houses than faces

Question 18

What are the two views of visual perception (background, Egner & Summerfield)?

Answer 18

1. Predictive coding (perception is inference, ie, conclusion based on reasoning from the data; each level of visual hierarchy has representation and error units)
2. Feature detection (visual neurons just respond to features of an object) –> alternative view in paper

Question 19

What is the research question and their hypothesis (Egner & Summerfield)?

Answer 19

RQ: Does BOLD activity in the FFA reflect responses to expectation and surprise, or just face features?

General H (predictive coding): FFA responses to faces and houses should be most different when face expectation is low

Alternative H (feature detection): there will always be more FFA activation to faces than to houses, regardless of expectation

Question 20

What were the IV and DV (Egner & Summerfield)?

Answer 20

IV: stimulus probability (% of time, face vs. house); stimulus feature (face/house); target vs. nontarget

DV: reaction time. BOLD response in FFA and PPA

Question 21

How did they manipulate expectation (Egner & Summerfield)?

Answer 21

Participants were told that different coloured frames meant different likelihoods that a face would appear

Question 22

How did they keep attentional demands constant across conditions to avoid differences in attention being a confound?

Answer 22

They included upside-down faces (allows response to be to what you actually see rather than what you expect to see)

Question 23

What condition represents the error signal in Egner & Summerfield?

Answer 23

FFA activation reflecting surprise - surprise requires error signal to be sent up for adjustment

Question 24

Did Egner & Summerfield find that expectation and surprise contribute equally to the FFA population response?

Answer 24

In the PC model that fit best, surprise contributed 2x as much as expectation

Question 25

What were the conclusions (Egner & Summerfield)?

Answer 25

Prediction coding models describe the process of visual inference better than feature detection models; encoding prediction and error is a general characteristic of how the brain works Results are incompatible with feature detection model - We know more about the relative contributions of prediction and error units (error units have higher contribution than representation units) - We don't know how much the BOLD response reflects top down vs bottom up inputs - There may be an interaction with attention

Question 26

What do feature detection models state?

Answer 26

Sensory information is transmitted in a hierarchical feedforward way; focus is on features of objects

Question 27

How do feature detection models and perceptive coding models differ?

Answer 27

FDM: Sensory information is transmitted in a hierarchical feedforward way; focus is on features of objects - bottom-up PDC: The brain continually generates models of the world based on context and information from memory to predict sensory input - bottom-up and top-down are compared