Artificial Intelligence and Neurocognition
Lecture 1 – Introduction & History of AI
What is artificial intelligence?
The study of the computations that make it possible to perceive, reason, and act
The study of how to build or program computers to enable them to do what minds
can do
o AI draws from disciplines of psychology (but puts greater emphasis on
computation), on computer science (but puts greater emphasis on
perception, reasoning, and action), and on neurophysiology
o Psychology is an inverse problem AI uses forward modeling: design a
system and observe how it behaves
How did the field of AI develop?
Physical brain mental mind
o Descartes: dualism; materialists: mental states = physical states
1940s: McCulloch and Pitts’ three principles:
1. Basic physiology
2. Propositional logic
3. Turing’s theory of computation
o Any computable function can be computed by a network of neurons
o All logical operators can be implemented by simple neural networks
Weak vs. strong AI
Weak AI:
1950s: Turing test (imitation game) non-sentient AI
o Intelligent machine = if we cannot distinguish it from a human in
conversation
o Makes no claims about the underlying mechanisms
o Intelligence is determined by complex grammatical structures &
realistic world knowledge
Opponent: Searle
o A collection of cells can lead to thought/action/consciousness
consciousness requires actual physical-chemical properties of actual
human brains only brains cause minds
o Chinese Room argument = one (non-Chinese speaking) person alone
in a room with a book that contains every possible answer (in
Chinese) to questions that are being asked from the outside (in
Chinese) answers will be correct, but does the person really
understand? Can computers really understand?
o Rule-based manipulation of symbols does not constitute intelligence
o Machine behavior may appear intelligent, but it does not reflect true
intelligence or sentience
Strong AI:
o Believes that intelligent systems can actually think
o Assumes that the human mind is an information processing system, and that
thinking is a form of computing (cognitive psychology)
, o Does an accurately enough simulated human mind have all the same
properties as an actual human mind?
o 1950s: Minsky and Edmonds’ SNARC first neural network computer with
40 neurons
o 50/60s: Dartmouth Conferences coined the term artificial intelligence
o 50/60s: general problem solving using a physical symbol system computers
playing checkers, invention of Lisp (high-level AI language)
o Symbolic AI (GOFAI) does not concern itself with neurophysiology
o Human thinking is a kind of symbol manipulation: IF (A > B) AND (B > C) THEN
(A > C)
o Knowledge-based/expert systems
o ELIZA = an early natural language processor (Weizenbaum, 1965). Used
simple techniques to create the illusion of understanding
o Anthropopmorphization of computers is a ‘trick’
o STRIPS = an automated planner (Stanford Research Institute Problem Solver).
Divides a task into subgoals, identifies necessary actions
Susceptible to the Sussman anomaly: subgoals are limited in
solving the final goal
o MYCIN (expert system) = emulates the decision-making ability of a human
expert. Designed to diagnose and recommend treatment for blood infections
based on simple if-then rules with certainty factors.
Reached an accuracy of ~69%, which was better than
physicians never used due to ethical and legal difficulties
60/70s: overconfidence in AI systems not as powerful as many thought AI
winters (no funded research) symbolic AI does not suffice
1980s: Rumelhart & McClelland: the PDP research group connectionism (=
parallel distributed processing, = artificial neural networks)
o Model of human memory, content-addressable
o Memory is stored in connections (synapses) between neurons inhibitory &
excitatory connections
o Connectionist AI = biologically inspired (based on structure of the human
brain); lesion tolerant (damaged networks can still process information);
capable of generalization (capable of learning)
o Neurons receive input through dendrites neurons send output through
axon highly connected (1000s of synapses) 20x109 neocortical neurons,
15x1013 cortical synapses computation is massively parallel (efficient)
o Neurons output a signal based on their input signal
o Multi-layer perceptrons are able to implement all logical operators, such as
AND, OR, XOR (not OR)
o Mental states are represented as N-dimensional vectors of numeric
activation values over neural network units
o Memory is created by modifying the connection strength (weight) between
units
o Connectionist AI can solve complex, non-linear or chaotic classification
problems
o No a priori assumption about problem space or statistical distribution
, o Artificial neural networks can compute any computable function (McCulloch
& Pitts)
o Mainly used for pattern recognition
80/90s: ‘modern AI’ first time a computer (IBM’s Deep Blue) beat a grandmaster
at chess (1997); last time a human beat a top chess computer (2005); smartphone
running chess software equals Deep Blue’s performance (2009)
2000s – present: data mining offers huge quantities of data; deep learning offers
representation at many levels; Bayesian networks deal with uncertain knowledge;
deep reinforcement learning can learn to act from rich, noisy data
Deep networks: adding more layers adds to dimensionality of classification
o Multiple representations offer multiple levels of abstraction
o Recurrent connections can maintain context, temporal information
o Combination is hot topic: Google is investigating motion classification and
content classification
Learning
If we don’t want to preprogram all knowledge, systems should be able to learn
A computer program is said to learn from experience E with respect to some class of
tasks T and performance measure P, if its performance at tasks in T, as measured by
P, improves with experience E
Machine learning
o Supervised learning = external knowledgeable supervisor presents the system
with correctly labeled training data
o Unsupervised learning = discover hidden structure in data without labeled
data
o Reinforcement learning = learning from a feedback signal
Classification
o Determining group membership based on input data
Does this MRI image show a brain tumor?
Regression
o Predict outcome data based on input data
Given its location, surface area, and number of rooms, can we predict
the value of this house?
Conclusion
Philosophical implications
o Weak AI: machines can simulate human intelligence using clever tricks
o Strong AI: a well-programmed machine that exactly emulates the human
brain is a mind, and thereby intelligent
Approaches to AI:
o Symbolic AI: intelligent behavior through manipulation of symbols
o Connectionist AI: representations in the brain are distributed, processing
massively parallel
Lecture 1 – Introduction & History of AI
What is artificial intelligence?
The study of the computations that make it possible to perceive, reason, and act
The study of how to build or program computers to enable them to do what minds
can do
o AI draws from disciplines of psychology (but puts greater emphasis on
computation), on computer science (but puts greater emphasis on
perception, reasoning, and action), and on neurophysiology
o Psychology is an inverse problem AI uses forward modeling: design a
system and observe how it behaves
How did the field of AI develop?
Physical brain mental mind
o Descartes: dualism; materialists: mental states = physical states
1940s: McCulloch and Pitts’ three principles:
1. Basic physiology
2. Propositional logic
3. Turing’s theory of computation
o Any computable function can be computed by a network of neurons
o All logical operators can be implemented by simple neural networks
Weak vs. strong AI
Weak AI:
1950s: Turing test (imitation game) non-sentient AI
o Intelligent machine = if we cannot distinguish it from a human in
conversation
o Makes no claims about the underlying mechanisms
o Intelligence is determined by complex grammatical structures &
realistic world knowledge
Opponent: Searle
o A collection of cells can lead to thought/action/consciousness
consciousness requires actual physical-chemical properties of actual
human brains only brains cause minds
o Chinese Room argument = one (non-Chinese speaking) person alone
in a room with a book that contains every possible answer (in
Chinese) to questions that are being asked from the outside (in
Chinese) answers will be correct, but does the person really
understand? Can computers really understand?
o Rule-based manipulation of symbols does not constitute intelligence
o Machine behavior may appear intelligent, but it does not reflect true
intelligence or sentience
Strong AI:
o Believes that intelligent systems can actually think
o Assumes that the human mind is an information processing system, and that
thinking is a form of computing (cognitive psychology)
, o Does an accurately enough simulated human mind have all the same
properties as an actual human mind?
o 1950s: Minsky and Edmonds’ SNARC first neural network computer with
40 neurons
o 50/60s: Dartmouth Conferences coined the term artificial intelligence
o 50/60s: general problem solving using a physical symbol system computers
playing checkers, invention of Lisp (high-level AI language)
o Symbolic AI (GOFAI) does not concern itself with neurophysiology
o Human thinking is a kind of symbol manipulation: IF (A > B) AND (B > C) THEN
(A > C)
o Knowledge-based/expert systems
o ELIZA = an early natural language processor (Weizenbaum, 1965). Used
simple techniques to create the illusion of understanding
o Anthropopmorphization of computers is a ‘trick’
o STRIPS = an automated planner (Stanford Research Institute Problem Solver).
Divides a task into subgoals, identifies necessary actions
Susceptible to the Sussman anomaly: subgoals are limited in
solving the final goal
o MYCIN (expert system) = emulates the decision-making ability of a human
expert. Designed to diagnose and recommend treatment for blood infections
based on simple if-then rules with certainty factors.
Reached an accuracy of ~69%, which was better than
physicians never used due to ethical and legal difficulties
60/70s: overconfidence in AI systems not as powerful as many thought AI
winters (no funded research) symbolic AI does not suffice
1980s: Rumelhart & McClelland: the PDP research group connectionism (=
parallel distributed processing, = artificial neural networks)
o Model of human memory, content-addressable
o Memory is stored in connections (synapses) between neurons inhibitory &
excitatory connections
o Connectionist AI = biologically inspired (based on structure of the human
brain); lesion tolerant (damaged networks can still process information);
capable of generalization (capable of learning)
o Neurons receive input through dendrites neurons send output through
axon highly connected (1000s of synapses) 20x109 neocortical neurons,
15x1013 cortical synapses computation is massively parallel (efficient)
o Neurons output a signal based on their input signal
o Multi-layer perceptrons are able to implement all logical operators, such as
AND, OR, XOR (not OR)
o Mental states are represented as N-dimensional vectors of numeric
activation values over neural network units
o Memory is created by modifying the connection strength (weight) between
units
o Connectionist AI can solve complex, non-linear or chaotic classification
problems
o No a priori assumption about problem space or statistical distribution
, o Artificial neural networks can compute any computable function (McCulloch
& Pitts)
o Mainly used for pattern recognition
80/90s: ‘modern AI’ first time a computer (IBM’s Deep Blue) beat a grandmaster
at chess (1997); last time a human beat a top chess computer (2005); smartphone
running chess software equals Deep Blue’s performance (2009)
2000s – present: data mining offers huge quantities of data; deep learning offers
representation at many levels; Bayesian networks deal with uncertain knowledge;
deep reinforcement learning can learn to act from rich, noisy data
Deep networks: adding more layers adds to dimensionality of classification
o Multiple representations offer multiple levels of abstraction
o Recurrent connections can maintain context, temporal information
o Combination is hot topic: Google is investigating motion classification and
content classification
Learning
If we don’t want to preprogram all knowledge, systems should be able to learn
A computer program is said to learn from experience E with respect to some class of
tasks T and performance measure P, if its performance at tasks in T, as measured by
P, improves with experience E
Machine learning
o Supervised learning = external knowledgeable supervisor presents the system
with correctly labeled training data
o Unsupervised learning = discover hidden structure in data without labeled
data
o Reinforcement learning = learning from a feedback signal
Classification
o Determining group membership based on input data
Does this MRI image show a brain tumor?
Regression
o Predict outcome data based on input data
Given its location, surface area, and number of rooms, can we predict
the value of this house?
Conclusion
Philosophical implications
o Weak AI: machines can simulate human intelligence using clever tricks
o Strong AI: a well-programmed machine that exactly emulates the human
brain is a mind, and thereby intelligent
Approaches to AI:
o Symbolic AI: intelligent behavior through manipulation of symbols
o Connectionist AI: representations in the brain are distributed, processing
massively parallel
