Reading...
Play button
Play button
Use LEFT and RIGHT arrow keys to navigate between flashcards;
Use UP and DOWN arrow keys to flip the card;
H to show hint;
A reads text to speech;
64 Cards in this Set
- Front
- Back
|
good properties of a definition (necessary and sufficient)
|
necessary: everything that is a cat matches the def'n
sufficient: everything that matches the def'n is a cat |
|
|
the classical view of concepts
|
logical definitions
|
|
|
Problems for the classical view
|
- good def'ns are hard to find
- borderline cases: is a lamp or rug furniture? EX: bachelor = unmarried adult male, except there are many examples of men who match this def'n but who we don't consider to be bachelors (e.g. priest, life partner, 17 y.o.) |
|
|
typicality effects
|
typicality influences reaction time, generalization.
EX: something that looks a little like a bird that we have never seen before. it's hard for us to say if it's a bird instantly |
|
|
prototype and exemplar theories
|
concepts have a prototype, typicality is a function of similarity to the prototype. but how do we know info about a prototype's variability?
exemplar theory: we store every instance that we see in memory -> exemplar theory supports prototype theory |
|
|
The Wason task, effects of context, and possible explanations
|
logic puzzle that tests human intuition; asks testtaker to identify which cards must be turned over to prove a certain rule. many failed. however, if we ask them to prove a rule that they know better (EX: if someone is drinking, they are over 21), it's an easy puzzle
|
|
|
The basic idea behind the heuristics and biases approach (a question regarding the relationship between them)
|
if people use heuristics to solve a problem, what biases influence their decisions?
|
|
|
The representativeness heuristic and how it can lead people astray
|
- if process A is similar to event B, where process A yields event B, then the probabilities between them are judged to be high
- if we can generate more examples of A from memory, A will appear to be more likely |
|
|
main points about human reasoning
|
we act in ways other than according to probability theory. what we remember is not simply based on frequency
humans don't seem to categorize things using rules, and human reasoning is not in perfect accordance with probability theory |
|
|
the idea behind connectionism
|
theory of how thought works uses spreading activation in an artificial neuron network
|
|
|
how ANNs work
|
uses a network of many concepts, analogous to neurons since there are so many of them. hypothesis is that nodes work similarly to neurons (firing above a certain threshold)
|
|
|
localist vs. distributed representations
|
localist: one node per concept
distributed: a stream of activated nodes per concept |
|
|
properties of distributed representations
|
we can plot these representations on a 2D plane and find a line that indicates a boundary of recognizing
noise, interpolation (70% dog) —> explains how humans can hear and identify things in loud situations, and how if something sounds like a dog and a cat, we can guess what percent "dog" it is. it's not just "dog" or "cat". |
|
|
associative learning / Hebb rule
|
neurons that fire together wire together. activations that happen often will continue to happen because weights between them become stronger each time.
|
|
|
kinds of learning problems: supervised and unsupervised
|
unsupervised: we learn by picking up on statistical regularities. like classical conditioning
supervised: a "teacher" tells us if something is right or wrong, FEEDBACK. like operant conditioning |
|
|
idea behind error-driven learning (hillclimbing, gradient descent)
|
recalcuate weights so that error is minimized
if we have a differentiable function of error, we can find minima. these points are where error is lowest. |
|
|
problems with hillclimbing algorithms
|
we might find many local minima when we're acutally looking for the absolute minimum. because of this, we might be making a mistake but think it's correct since we're at minimum error
|
|
|
perceptrons and their problems
|
feed-forward neural network involving input/output layers
- by this theory, we can only learn linearly separable categories (no XOR function) e.g. either dog or cat |
|
|
basic idea behind backpropogation
|
we compute the error, use gradient descent, propagate error to a hidden layer, use gradient descent again
- no cycles or feedback, only forward |
|
|
arguments for and against connectionism
|
distributed representations are key to understanding thought
- no one is really sure where representations come from in the first place artificial neural networks are attractive compared to logic — still not really sure how neurons work though explains "fuzziness" that humans have in reasoning — but symbols have gotten us very far the idea that there is a hidden layer is telling of how much we know similar to Nativist vs. Empiricist |
|
|
how AI fits into cogsci (linguistics, psychology, neuro)
|
- allows us to explore principles that make the mind work
- we can test our theories |
|
|
analysis by synthesis
|
in many sciences we break things down into the most elemntary pieces. but when we don't know enough about the mind to begin with, it's more helpful to start from scratch and see what elementary pieces are necessary to help our theory develop
|
|
|
the Turing test and its flaws
|
test does not necessarily measure intelligence, but rather ability to act intelligently
- lacks creativity, mistakes, ESP (extrasensory perception) |
|
|
strong and weak AI
|
weak: act intelligently
strong: actually think like a human |
|
|
basic idea behind the General Problem Solver: means-ends analysis
|
goal-oriented: every step it takes is the best step to get it closer to its goal ("hill-climbing procedure").
- program is based on human problem-solving strategies (heuristics) |
|
|
How early AI systems worked (programs, symbols, operations)
|
designed programs that used symbolic representation. programs could recognize a symbol, and then perform an operation
|
|
|
problems for early AI
|
- uncertainty
- symbol grounding problem - to think like a human, there would be an infinite amount of data to work with from the world--too many algorithms to write! |
|
|
how psych fits into cogsci
|
through psych, we can better understand people's representations. a lab setting lets us isolate variables and measure properties of the mind
|
|
|
how should we investigate cognitive architectures?
|
we should use the idea of symbolic representation to understand the architecture of the mind
|
|
|
modularity hypothesis
|
different modules for individual processes (DOMAIN SPECIFIC) such as vision, auditory, etc.
modules are INFORMATIONALLY ENCAPSULATED, each are opaque to one another, no communication between modules EX: if we see a visual illusion, we'll still see the illusion even if we know it's an illusion |
|
|
evidence in favor of a "language of thought"
|
- mind uses propositions
- we can recall semantic information, but not syntactic |
|
|
semantic networks and spreading activation
|
- many propositions linked to one another. the more related one proposition is to another thought, the fewer arrows it must follow to reach that proposition node
- any node can be "activated" and this activation spreads. this can explain "train of thought" |
|
|
evidence for semantic networks (response times)
|
studies have shown how related concepts can be used to predict response times for making decisions
|
|
|
scripts and schemas
|
schema: list of properties in a semantic network that surround a certain concept, work on a general level
script: representations of a sequence of actions associated with a type of activity (have to do with roles, goals of a certain situation) |
|
|
challenges for view of the mind as a symbolic system
|
similar to early AI: we can have rules (propositions, scripts, schemas), but our mind is still subject to:
- symbol grounding - uncertainty - also unclear how this information got here in the first place |
|
|
mental rotation
|
an experiment to test mental imagery: degree of rotation was proportional to response time in identifying if object was rotated correctly
|
|
|
key claim in the mental imagery debate
|
claim: mental imagery has its own system of representations/operations
|
|
|
positions in the mental imagery debate
|
Kosslyn: experiments show that response time is linearly related to distance (scanning maps in this case)
--> we have a system of representations specifically for imagery Pylyshyn: we decompose images into mathematical propositions |
|
|
how neuroscience fits into cogsci
|
explores the hardware of the human mind, the mechanisms that do the imformation processing
|
|
|
different views of the brain (classical. connectionist)
|
phrenology: different parts of the brain are responsible for different states of mind and concepts(anger, sadness, hope)
decartes: all perceptions went to the pineal gland, but this turned out to be false connectionist: ANN |
|
|
corpus callosum
|
connects brain hemispheres
|
|
|
diencephalon
|
thalamus, pituitary, pineal
|
|
|
midbrain
|
primitive sensory/motor
|
|
|
hindbrain (medulla, pons, cerebellum, spinal cord)
|
balance, posture
|
|
|
occipital
|
vision
|
|
|
temporal
|
auditory, visual info (FFA)
|
|
|
parietal
|
computation, attention, integration of sensory information
|
|
|
frontal
|
motor controls, planning, speech, higher-level cognition
|
|
|
Cerebral asymmetries and how the brain connects up to the world (contralateral links)
|
right-handed person has more language processing done in left side of brain than left-handed person. the brain has contralateral links: right hemisphere receives left eye visual info, control left side of body, vice versa
|
|
|
neuron parts
|
nucleus: houses the working parts of the cell
dendrite: carries signals axon hillock: where signal begins axon terminal: where signal ends |
|
|
how a neuron works (change of charge)
|
neuron has a charge of -70 mV, filled with negatively charged ions. chemical signals from other cells modify this charge, and then there's a chemical reaction between neurons (the signal is passed on)
action potential releases neurotransmitters along synapse which connects axon terminal of one neuron to the dendrite of another neuron |
|
|
learning in neurons (wire together, fire together)
|
if an action potential is repeatedly released, it will will begin to release a greater potential. so if we do something repeatedly, the associated neurons will fire more
|
|
|
why do we study neural representation?
|
understanding the representations in our brain can help us in understanding those in our mind
- it would be helpful to understand how neural representations are ACQUIRED |
|
|
methods for studying neural representation
|
use animals: PUT DYE IN A CELL, LESIONING
put dye into a cell that is firing to see where its signal goes lesioning (killing) part of the brain and looking at the consequences - in humans: brain-damaged patients |
|
|
neural representation in visual cortex
|
simple cells: they see lines
complex cells: see angled lined hypercomplex: see corners, terminated lines |
|
|
topographical maps in the brain
|
parts of the cortex respond to info that's being trasmitted in the same dimension (neighboring)
|
|
|
scomata
|
where optic nerve meets the retina, forms a blindspot, but our brain "fills in" the visual info
|
|
|
blindsight
|
the links between the eyes and the brain are damaged, but the eyes work fine. some information from eyes goes to other parts of the brain, so blindsighted people can "see" a little, through subconscious
|
|
|
EEG (electroencephalography) imaging
|
brainwaves
|
|
|
fMRI (functional magnetic resonance imaging)
|
bloodflow
|
|
|
PET (positron emission tomography)
|
insert radioactive dye and see where it goes
|
|
|
fusiform face area and prosopagnosia
|
evidence of modularity hypothesis?
people with damaged to the FFA has trouble recognizing faces (prosopagnosia), even their own |
|
|
neural plasticity
|
using certain parts of the body may increase representativeness in brain, decreases with time
|
|
|
localist and distributed nodes in neural representations
|
localist: "grandmother cell" -- fires only for your grandmother
distributed: we can predict movement averaging the directions in which it moves, and weight these measurements by the amount certain neurons fire |
