lecture 1: introduction + pp Flashcards
1
Q
active vs. passive perceiver
A
- much of (changes in) our sensory world is driven by our own actions.
- one of the major computations that our brains execute are therefore to interpret the sensory world as we interact with and change it.
- however, we study the brain as a passive recipient of information by studying it in a restricted setting. this gives us a lot of information, but might not be ecologically valid.
2
Q
top-down vs. bottom-up connections in V1
A
- from V1 to V2, there are nearly 10x more feedback (top-down) connections than feedforward (bottom-up) connections.
- implies that visual processing in V1 is heavily influenced by memory and predictions (top-down), rather than being purely reactive
3
Q
boundaries between perception, cognition, and action
A
- very blurry
- there is a lot of overlap in brain areas labeled with cognitive functions such as selective attention, planning, and decision making
4
Q
embodied vs. isolated
A
- we do not experience the world with our bodies, but because we have bodies
- evolution: our cognitive functions have evolved alongside bodily systems, showing that cognition is not just a brain-based process, but is deeply rooted in the entire body
- visceral signals don’t just react to the brain but actively shape how it processes information and experiences
5
Q
uncertainty of perception
A
- the brain only has direct access to its own neural activity, not the external world. the causes of the sensory signals (what’s really out there) are hidden from the brain.
- the brain must understand these causes to interact effectively with its environment for survival.
- the brain learns about the world through action and the sensory feedback it receives, adapting its understanding over time.
6
Q
DNNs and LLMs
A
- passively fed large amounts of training data, making them good at pattern recognition.
- limited to their training material, meaning that they lack transfer of learning, leading to mistakes humans wouldn’t make
- passive learning (bottom-up learning) leads to artificial understanding of information
7
Q
lack of overarching theory
A
there is no cohesive framework that ties together functions related to behavior, perception, cognition, and action.
8
Q
predictive processing theory
A
- potential unified brain theory
- free energy principle: the idea that all living systems (including the brain) strive to reduce the discrepancy between prediction and reality to maintain a stable state, which is key to its functioning and survival.
- rooted in physics and mathematics
9
Q
free energy minimization: perception and action
A
- perception and action play complementary roles in the minimization of free energy
- Perception: The brain tries to align its predictions about the world with the actual sensory data. This minimizes the gap between what the brain expects and what it perceives from the environment. (perceptual inference)
- Action: The brain can also act on the world to change its observations, gathering new sensory evidence to maximise evidence. By interacting with the environment, it helps reduce uncertainty about what’s happening around it. (active inference)
10
Q
basic imperative of any living organism
A
- Entropy and Boundaries: Organisms maintain a boundary between themselves and the world to resist the second law of thermodynamics, which drives systems toward increasing entropy over time.
- Survival through Predictions: To survive, organisms make predictions across timescales to act in ways that minimize entropy. When an organism’s expectations about its environment or internal states are violated, prediction errors occur, creating free energy.
- Minimizing Free Energy: The primary aim of organisms, including the brain, is to minimize free energy or prediction errors. It predicts to be in viable states. This ensures survival by keeping physiological variables stable.
11
Q
three ways of prediction error minimization
A
- active inference: change sensory input to align with the brain’s prediction through action (model confirmation)
- perceptiual inference: revising predictions - change expectations about what is sampled (model revision)
- precision weighting: mechanism to control the relative influence of top-down predictions vs. bottom-up input in a context sensitive manner
12
Q
predictive coding hierarchy
A
- hypothesis about how the brain might implement perceptual inference
- predictions about the world are sent down (descending predictions) and errors in those predictions are sent up (ascending prediction errors).
- Hierarchy: Lower layers focus on processing detailed features and comparing them to predictions. higher layers help refine those predictions (more slowly) based on accumulated knowledge and experience.
13
Q
active inference
A
- actions of agents change the sensory input to align with the brain’s predictions through action (model confirmation)
- also provides the brain with a method to check if its models are correct (model testing): If I sample the external environment in this way, this observation (sensory state) should result.
- Thus, action is critical for building reliable models of external affordances
14
Q
active inference: planning of a sequence of actions
A
- done based how free energy is expected to be minimised in the future, as a function of discrete actions.
- this requires generative models with temporal depth.
15
Q
active inference: generative model and generative process
A
- generative process: our worlds evolve according to a dynamical process that generates observations (y) from hidden states (x*)
- generative model: our internal models accounts for observations (y) in terms of hypothetical hidden states (x)
- our inferences about these states based upon our observations then drive actions (u) that intervene on the processes generating our sensations
