1/35
Pages 30 - 37
Name | Mastery | Learn | Test | Matching | Spaced | Call with Kai | Chat |
|---|
No analytics yet
Send a link to your students to track their progress
Large language models
Software takes a sequence of words (tokens) and predicts the next word. The prediction is added to the input (autoregression) and the next word is predicted
Attention layers
Consists of multiple attention heads lie at the heart of the transformer architecture. Similar to convolutional layers and filters in image-recognition networks.
Deterministic prediction
Randomness is added post-hoc
Training steps
First step text prediction. Use text as fragment input, mask word and predict next word. Use backpropagation to update the network. Second step reinforcement learning. Have different predictions rated by humans.
LLM knowledge
Long-term (semantic) memory. Boundaries are fuzzy, can lead to confabulations or hallucinations.
LLM prompts
Working memory. Relevant information can be inserted into the conversation (possible hidden)
Innate language
Chomsky (and Fodor). Training neural networks (to predict text) will uncover grammatical structure. Long-range structures of text.
LM
System for haphazardly stitching together sequences of linguistic forms it observed in its vast training data. It is a stochastic parrot, they combine without any reference to meaning
Humans learn concept not statistical models Mitchell & Krakauer 2023
Abstract knowledge and experience to make robust predictions (generalizations and analogies), reason compositionally and counterfactually, intervene on the world (to tes hypothesis) and explain ones understanding
Multiple realizability
Just because two systems have the same output does not mean they work in the same way
Psychology of LLMs
To what extent LLMs are models of human cognition. We can not look inside (proprietary). We test if LLMs have faculty X
Wason selection task
Prompt: “You are shown a set of four cards placed on a table, each of which has a number on one side and a letter of the other side. The visible faces of the cards show A, K, 4, 7.
Q: Which cards must you turn over in order to test the truth of the proposition that if a card shows a vowel on one face then its opposite face shows an even number.
Training of LLMs
Orders of magnitude more data than humans development
Ethics
Financial and environmental cost, efficiency should be an evaluation metric.
Dimensions of consciousness
Wakefulness (state), gradual and varies over time. Awareness (experience), to be conscious. Normally they correlate
Levels of wakefulness
Coma, no reactions and no sleep/wake cycles. Unresponsive wakefulness syndrome (vegetative state), sleep/wake cycle and autonomous reactions. Minimally conscious state, UWS with some non-reflex movements. Locked-in syndrome, fully awake and conscious but unable to move.
Perspectives on cognitive agents
Third-person perspective, work backwards from observable behaviors and capacities to information-processing mechanisms that could generate these.
First-person perspective, target of investigation is a rich conscious life full of feelings, emotions etc.
Knowledge argument
“Mary is confined to a black and white room is educated through black and white books and through lectures relayed on black and white television. In this way she knows all the physical facts about us and our environment in a wide sense of physical which includes everything in completed physics, chemistry and neurophysiology. It seems that Mary does not know all there is to know. For when she is let out of the black and white room or given a color television, she will learn what it is to see something red.”
Congruent
Category of prime == category of target
Incongruent
Category of prime ≠ category of target
Dissociations
Dissociation, in one type of the brain damage we have ability A normally while ability B is impaired then A and. Bare dissociated
Double dissociation, in disorder we have ability A normally and B impaired and second disorder B normally and A impaired
Neglect
Lesions in the right parietal and frontal lobes, patients lack awareness of sensory events on the contralesional side
Blindsight
Lesions in the visual cortex, patients are aware of this
Interim conclusion
Types of information processing and problem solving that can take place without consciousness and compare them with those that require consciousness
Milner & Goodale’s theory
Two pathways, dorsal pathway (where) and ventral pathway (what). Vision for acton (dorsal) deals with how actions are executed. Vision for perception (ventral) deals with initiating deliberate action
Conscious condition
Activation in dorsal and ventral stream
Nonconscious condition
Robust levels of activity in dorsal stream but not ventral
Priming
Non-conscious information is short-lived. Conscious information has a larger impact.
Types of consciousness Ned Block
Phenomenal consciousness (P-consciousness) is raw experience (qualia). Experiential properties of sensations, feelings etc.
Access consciousness (A-consciousness) is reportable info on perception, introspection and memory. States of direct control of thought, reasoning etc.
Chalmers
No single problem of consciousness. Distinction between cluster of easy problems and one difficult problem (hard problem)
Easy problems
How an organisms ability to discriminate, categorize etc. How a cognitive systems integrates information. How and why mental states are reportable. How a cognitive system can access its own internal states. How attention gets focused. The deliberate control of behaviour. Difference between wakefulness and sleep
Theories of consciousness
How can we explain the difference between mental states that are conscious and that are not. Higher-order perception, mental states are conscious when they are the objects inner sense (introspection) or quasi-perceptual faculty
Global workspace theory
Attention spotlights the contents of conscious experience. Rest is dark and unconscious, competing to gain access. Conscious experience is limited by attention, a gate-keeper and amplifier
Functions
Intentional control of actions, durable and explicit information and ability to plan new task through combining mental operations
Approaches to consciousness
Cognitive scientists study different aspects of conscious mind and how consciousness contributes to action and cognition.
Mysterians say consciousness is a mystery that we have no idea how to tackle with cognitive science
Vitalism
Biology and chemistry are incapable of explaining the difference between living and non-living so there must be something else (elan vital, vital force)