5-10 Flashcards
Top Down / Bottom Up
Top:start by identifying the way the system solves problems, and reps it uses
bottom: start at nueral(implementation level and work your way up)
Reward Learning
understanding, prediction and responding to positive negative outcomes…. in humans DA and other NT’s. but think abou it can fail.
in AI reward learning is critical reinforcement learing. Reward signals agents to update policy
Reverse Engineerin nuerons
Donkey Kong MOS6502. We don’t lack in data, but we lack in interpreting the data.
We Lack a theory. An exapanded set of tools and analyses need to be developeed
Mind- Brain theory
The mind is what the briain does.
knowledge chucnk moves in working memory and directs procedual knowledge
Knowledge correction is the path to being less wrong
an ascpect of cognition is symbol manipulation
symbols represet the world “out there” by moving our body according to them we interact with the world…. To precieve somethign means we have a reprsentation inside our mind
Mental Reps are like deskop icons
symbols are not reality. They stand for reality…. they allow us to react with reality.
a functionalist approach to knowledge helps to futher reavel symbols/ideas/concepts isn’t TRUTH…. but instead usefullness.
Semantic network
grows and grows.
Ebbinghaus
Forgetting curve.
Unification Challenge
the need for unifcation – Alan Newell, - data was fragmenting rather than converging.
endless pappers without progress.
you can’t play 20 questions with nature and win.
to arrive at cognition we need theories to integrate data
Dustbowl empiricism
science approach that makes empirical obs and collects data rather than establishing theoretical framework.
Parsimony
basic to all science. Choose the simples scientific explantion
Lakatosian framework
integrate data under unifying explanation
progressive research are centred on the notion of empir knowledge foundation, where new theories and methods lead to novel factual discoveries
Newell and cog architecture
depict all the invivisble rooms of the mind.
UTC is a single set of mechansism that account for all of cognition - motor control, perception, memory and problem solving.
UTC must explain
how intel orgs respond to enviro
how we exhibit goal-directed behaviour and choose goals rationality
how we use symbols
how we learn from experience
how the mind exits as a single system with no supervisor
2 basic assumptions of utc
1the mind is too complex to be understood by one discipline - needs to be interdiscplinary
2the mind has to be understood at many different levels - integrated multilevel explantion
UTC theorys to unify
perceptiual symbol sysems - barasalou
working memory - baddeley
memory via affordances - glenberg
long-term working memory - ericsson & kinstch
situated cognition - varela, thompson & roach
global workspace theory - baars
cognitive architecture - sloman
Marrs Tri level
computational - what, why
algorithmic - how
instanttition - physical/biological
way to understand levels of abstraction. - computational processes involved in information processing.
newell - implementation, computational, knowledge
2 models of information processing
pssh - turing, classical symbolic AI
connectionism/neural nets -
Process models - information processing approach
first meta theory in cognitive science - designed to explain cognition - process models depict how model reacts over time.
psychology process models
behaviourist - stimulus, black box, response
cognitive - input, mediation process, output
units of cognition
instantiating empirical data
-if-then (production)
-knowledge chunk
-module
-buffer (working memory)
Meta-knowledge
Represents cognition-
symbols are about themselves - a meta represnetionais about itself.
a sierpinski triangle
MK refers to knowledge about knowledge or cognition.
it represents properties of cognition
awareness or control of the mind - thinking about thinking.
motnitor - extrernal and interal mental state
Cybernetics
underlying idea to build mechanical models. Fundamental insight
Architecture / Model
arch - Multiple depicts general operations / a single set mechanisms for all processes
model - single depicts phenomena
ACT-R
adapative control of thought - rational
John R anderson
it represents cognition as a system of modules
declarative vs procedural.
procedural knowledge (computational) Production rules
specified as production rules
- transform infor, and change state of system to complete task
if - then rule….. or condition-action
production rules are outside of consciousness
DPO
direct procedure optimation
What makes a good cog model
- external (measurable) behaviour
- internal states and processes
compared through THEORY of VALIDATION
Alt Cog architectures
Classic info processing —psychology
BOXOLOGY (ignore the brain)
Eliminative Connectionsim
- more nureal nets
Mathmatical aproach
-ignore architecture
MMoM
Machine theory of the mind.
TOM - relates to processes an agent use to represented the mental states of itself and others.
Machines require MT to make inferences and predicitions about human cog states
ToM for human-machine teaming
robots can manipulate humans
social INSCRIPTION EXPERIMENT- pet rock
Chess and GO
move 37 - the alien move
deep convolutional nueral nets, reinforcement learning, principled (learning) alogorithms
chess - humans only care about shallow moves