Predictive Processing Part 1

What’s the traditional or passive image of perception?

The light hits your eyes, sensory information enters brain, processed, the perception begins with sensory input and works its way bottom-up

What is the “cognitive couch potato” view of the brain?

Your brain is waiting for stimulation and information, then processes information progressively inward from the sensory periphery

Instead of waiting for sensory information, what is your brain like in Predictive Processing?

Constantly trying to predict what sensory information it is about to receive and involves two directions of information

What’s the 1st direction of information in Predictive Processing?

Top-down, brain generates predictions about what is out there

What’s the 2nd direction of information in Predictive Processing?

Bottom-up, incoming sensory information tells the brain where its predictions were wrong

How is perception defined in Predictive Processing?

A combination of top-down, knowledge-based predictions and bottom-up sensory evidence

What does prediction and the actual sensory output help do in Predictive Processing?

Match one another reasonably, if not then prediction error arises

What does Prediction Error in Predictive Processing tell the system?

Something about the current prediction or model isn’t adequately explaining the sensory information then improves prediction

What is a Generative Model?

Stored probabilistic knowledge about the world in the form of a generative model

What is accumulated in the Generative Model?

What kinds of things exist in the world, how they tend to behave, and what sensory input those things would generate

What does Controlled Hallucination?

Means perception involves the brain generating a hypothesis about what is out there, controlled by sensory evidence

What’s the difference in Hallucination and Normal Perception?

Prediction dominates without being adequately constrained by sensory evidence, prediction is continually constrained/corrected by sensory evidence

What does the phrase “explaining away” mean in Controlled Hallucination?

The prediction successfully accounts for this part of the sensory signal, so there is no remaining mismatch that needs explaining

What’s 1. Predictive Processing?

Proposes that perception results from the brain using a generative model to generate top-down predictions about sensory input

What’s 2. Predictive Processing?

Incoming bottom-up sensory information is compared with those predictions, and mismatches produce prediction errors that drive updated predictions and learning

How does Predictive Processing get scaled up?

By predicting sensory information at different spatial and temporal scales, the system can learn about about things ranging from lines and edges to people, meanings, goals, intentions

Why a hierarchy in Hierarchical Predictive Processing?

The brain makes predictions makes predictions at multiple levels of complexity at the same time

What is the structure of Predictive Processing in a hierarchy?

Higher level predicts lower level, lower level compares prediction with actual input, mismatch becomes prediction error, error travels upward and higher-level hypothesis gets revised

Why is the hierarchy in Predictive Processing useful?

Because perception has to connect simple sensory features to complex interpretations

What idea in the hierarchy in Predictive Processing scaled up?

To predict visual input, the system learns specific things and can learn about things such as goals, people, faces and intentions

Who demonstrates an example of the hierarchical system and how it could be useful?

Rao and Ballard, 1999 and shows how the hierarchical predictive system could actually learn

What’d Rao and Ballard do?

Their artificial neural network was givens lots of image patches from natural scenes and tried to predict its own evolving sensory states

2. What’d Rao and Ballard do, the first-level?

After exposure to thousands of natural image patches, the first-level learned features like orientated edges and bars

3. What’d Rao and Ballard do, the second-level?

The second-level network learned more complex patterns involving larger spatial contexts

4. What’s the result of Rao and Ballard?

The model reproduced a known visual phenomenon called end-stopping

5. Rao and Ballard, what does end-stopping mean?

Some neurons respond strongly to a short line in their receptive field, but their firing decreases when the line becomes longer

6. How is the results of Rao and Ballard interpreted?

The stronger firing to short segments as error/mismatch because the higher level had predicted a longer line

1. What is the relevance of Rao and Ballard?

Shows how prediction can drive learning, the system develops useful representations simply by trying to predict its sensory input

2. What is the relevance of Rao and Ballard?

Shows how neural responses can potentially reflect prediction error, rather than just passive feature detection

3. Relevance, in a very different environment, end-stopped cells could learn very different responses, why is that important?

Because the system is sensitive to the statistical structure of the world it learns from

1. Connection, how does Predictive Processing connect to Connectionism?

Prediction errors can drive learning by adjusting connection weights through forms of gradient descent

1.5 Connection, what idea does connectionism propose?

Cognition emerges from networks of simple units whose connection weights change through learning

1. What does Hierarchical Predictive Processing explain?

Explains perception as a constant exchange between top-down predictions and bottom-up prediction errors across multiple levels of abstraction

What is the Bayesian inference and how does it relate to Predictive Processing?

A way of determining the probability of a hypothesis given the available evidence

What are the three pieces in the Bayesian Inference?

Prior, likelihood, and Posterior

1. What is Prior in Bayesian Inference?

Represents the probability assigned to a hypothesis based on previous knowledge or expectations

2. What is Likelihood in Bayesian Inference?

The compatibility of the sensory evidence with the particular hypothesis

3. What is Posterior in Bayesian Inference?

The updated belief given what was expected and just observed and can then effectively become a prior for future reference

How does Bayesian Inference fit in Predictive Processing?

How prior expectations and current sensory evidence are combined to arrive at perception

1. Connection, how do Bayesian Inference connect to Generative Models?

The model provides expectations/priors that allow the system to make predictions, then prediction error updates the model to become better over time

2. Connection, how could Predictive Processing relate and connect to Computationalism?

A particular image that might involve probabilistic inference over a generative model, using predictions and prediction errors to estimate the sensory input