Lesson ? - Revision for Assessment 1 Flashcards
When was the Gestation and Birth Period of AI
Lesson 3
1943-1956
When were the Golden Early Years
1956-1969
When was the First AI winter?
1966-73
When was the rise of Knowledge-based and Expert Systems?
1969-89
When was the start of the rise of New paradigms
continuing until now
1986
When was the start of the birth of the Scientific Method, Big Data and Deep Learning?
Continuing until the present day
1987
what was Pitts-McCulloch’s paper about
the first description of a neural network
Marvin Minsky was a student of who?
Pitts and McCulloch
Who built the first neural net machine?
Marvin Minsky and Dean Edmonds
name of McCulloch-Pitts 1943 paper
A Logical Calculus of Ideas Immanent in Nervous Activity
1st work in AI
McCulloch Pitts 1943, A Logical Calculus of Ideas Immament in Nervous Activity
Who is credited with creating/popularising formal propositional logic
Russell and Whitehead
neurons
electrically excitable cells that process and transmit info through electrical and chemical signals
3 parts of the artificial neuron
Dendrite, Soma, Axon
phi is a ______ function
step-like
what is fed into a neuron
inputs with associated weights
what does the neuron do with the inputs and their associated weights
computes the weighted sum and passes it through a NON-LINEAR TRANSFER FUNCTION
McCulloch and Pitts showed that
all logical connectives/any computable function could be computed by a neural network
Donald Hebb
proposed an updating rule for modifying connection weights in artificial neurons allowing a network to be trained
Hebbian Learning
when neurons fire together, they adapt and form stronger connections with each other through repeated use
name and date of first neural net computer
SNARC, 1951
Who built the SNARC
Marvin Minsky and Dean Edmonds
SNARC was used to
model the behaviour of a rat in a maze searching for food
neural nets + GPUs =
deep learning
the 40 neurons in SNARC were simulated by
3000 vacuum tubes
Logic Theorist written by
Herbert Simon, Allen Newell, JC Shaw
Logic Theorist was called the first…
AI Program
Logic Theorist was created when?
1955-6
why was Logic Theorist labelled as the 1st AI program
the first program deliberated engineered to mimic the problem solving skills of a human
Simon, Newell and Shaw proved how many of Russell/Whitehead’s theorems?
38
Name of Russell and Whitehead’s paper
Principia Mathematica
How many theorems where in Whitehead and Russell’s Principia Mathematica
52
what did Shaw/Herbert/Newell do with the final of Whitehead/Russell’s theorems
create a shorter version
axiom
self-evident truth
Logic Theorist:
Given some… (APT)
Axioms (A), theorems previously proved (P) and a candidate theorem to prove
how many methods were applied by the Logic Theorist
4
the methods applied by the Logic Theorist
substitution, detachment, chaining forward, chaining backward
what is substitution (Logic Theorist)
change logic expression (e.g. T) into a logically equivalent one (e.g. axiom in A) by substition of varaibles/replacements of connectives
detachment is also known as
modus ponens
detachment in Logic Theorist
create sub-goals to prove expressions (e.g. S and S -> T in order to prove T)
chaining forward in Logic Theorist
subgoals : A -> C then A -> B and B-> C
chaining backward in Logic Theorist
subgoals: A -> C then B-> C and A -> B
3 new concepts introduced using logic theorist
reasoning as search, heuristics and list processing
reasoning as search
proof viewed as a search starting from a hypothesis root node
explain the mechanism of proof as defined by “reasoning as search”
expand the proof along different branches according to deductive rules and stopping when the proposition to be proved is obtained
heuristics (as introduced by Logic Theorist)
proof tree grows exponentially thus some branches need to be pruned using rules of thumb/heuristics to limit search space while (hopefully) not losing the path to the solution
List processing
the list processing programming language IPL
IPL
the list processing language implemented by the authors in Logic Theorist
IPL serves as the basis for which language
LISP
who created LISP
John McCarthy
at non-root nodes, state = ?
{hypothesis + derived propositions}
new states are obtained from old staets by…
applying deductive rules
objective of Dartmouth Conference (quote)
every aspect of learning/any other feature of intelligence can be so precisely described that a machine can simulate it
The Darmouth Conference wanted to make an attempt to find out how to make a machine do what 4 things? (LAPI)
use language, form abstractions/concepts, solve problems reserved for humans, improve themselves
which 4 researchers proposed the Conference
McCarthy, Minksy, Rochester, Shannon (John, Marvin, Nathaniel, Claude)
the Conference was attended by who else?
Solomonoff, Selfridge, More, Samuel, Simon and Newell (Ray, Oliver, Trenchard, Arthur, Herbert, Allen)
what is considered the birth point of AI
the Dartmouth Conference
the GEY was marked by
overoptimism
which 3 computer scientists were particularly overoptimistic
Simon, Newell and Minsky
who funded labs and which did they fund?
ARPA (Advanced Research Projects Agency) - Stanford, MIT
GPS
General Problem Solver
who created General Problem Solver and when?
1959 - Simon and Newell
the GPS was meant to reason…
in a human-like way to solve any formalised symbolic problem
means end analysis - aspect 1
given a current state and goal state, attempt to reduce differenec between the two
means end analysis - aspect 2
operations (from current state) + their outputs -> creates subgoals to reduce distance to overall goal as much as possible
when did Newell and Simon formulate the physical symbol system hypothesis
1976
what led to Newell and Simon formulating the physical symbol system hypothesis
success of GPS and related programs as models of cognition
physical symbol system hypothesis
a physical symbol system has the NECESSARY and SUFFICIENT means for general intelligent action
physical symbol system
takes symbols, combines them into expressions, and manipulates them to produce new expressions
the physical symbol system implies that…
any intelligent system must operate by manipulating symbols
True/False: the implications of the physical symbol system are disputed
True
who made Geometry Theorem Prover?
Gelernter (1959)
Who made Advice Taker
McCarthy (1958)
which reasoned with general knowledge more broadly, Advice Taker or General Problem Solver?
Advice Taker
how could Advice Taker adapt to a new domain without reprogramming?
by separating explicit representation of world knowledge from deductive reasoning engine
who discovered Resolution Theorem Proving?
J.A Robinson 1965
what was resolution theorem proving?
complete theorem proving algorithm for 1st order logic
resolution theorem proving underlies which programming language?
PROLOG
Who/when created the Checkers programs?
Samuel
what did Samuel’s checkers programs demonstrate?
the potential of computers for non-numerical tasks/AI
what are the core ideas of alpha-beta pruning and minimax search strategy? as introduced by the Samuel’s checkers programs
search space too big to search exhaustively
what was the first machine learning program
Samuel’s checkers program
why was Samuel’s checkers program considered the first machine learning program
it learned from itself
what was the first computer game to reach the level of a respectable amateur?
checkers
why are so many programs limited to working in limited domains
unconstrained natural language is too difficult
who created STUDENT and when
Bobrow (1967)
who created SHRDLU and when?
T Winograd 1968-72
first chatbot & created when
eliza, 1964-6
who implemented PARRY and when
Kenneth Colby in 1972
which is more sophisticated? eliza or parry
parry
block worlds occurred primarily in which 3 domains
NLP, computer vision and robotics
SHAKEY : developed when and where
Stanford 1966-72
results of the SHAKEY project (3)
A* search algorithm, Hough transform (finding simple shape in images), Visibility graph method (used in robot motion planning)
SHAKEY
First general -purpose mobile robot to be able to reason about its own actions
SHAKEY used what to plan?
STRIPS
STRIPS requires what 3 things
intial state, goal state and actions
STRIPS develops…
a plan move from initital state to goal state by achieving appropriate subgoals
in 1976, the world’s fastest supercomptuter was capable of how many MIPS
1000
today, computer vision application require how many MIPs?
10^4 to 10^6 MIPS
combinatorial explosion
(N) (sum) (n=0) k^n where k is the number of potential moves and N is the number of steps ahead that needs to be examined
Moravec’s Paradox
high-level reasoning requires little computation effort but basic sensorimotor skills (e.g. perception/movement) are incredibly difficult to replicate
Who created the perceptron?
Frank Rosenblatt (1957)
the book Perceptrons was published by who when?
Minsky and Papert in 1969
The Frame Problem
actions change specific things but everything else - “the frame” remains the same
a single layer perceptron could not compute which function
the XOR
Single layer perceptrons could only solve….
linearly seperable classification problems
the qualification problem
the difficulty in specifying all possible pre-conditions for an action
US ALPAC Report 1966
withdrawal of funding
ALPAC purpose
appointed to evaluate progress in computational linguistics esp MT
UK Lighthill Report 1973
AI researchers had failed to address the issue of combinatorial explosion
US DARPA cancelled what at Carnegie Mellon University
Speech Understanding Research Programme
the US N______ R_______ ______ also cancelled funding
National Research Council
JR Lucas (1961) essentially reiterated
Turing’s anticipated Mathematical Objection
Godel Incompleteness Theorem
any formal system has true statements it cannot prove within the system itself
What did Lucas say about Godel’s Incompleteness Theorem?
a human can recognise the truth of these unprovable statements but a computer cannot
according to Russell and Norvig, Godel’s argument only applies to
TMs since computers are only approximations of TMs (no infinite memory)
Russell and Norvig’s 3 arguments to Lucas’ arguement
GIT only applies to systems powerful enough such as TMs, sentences which a given cannot consistently assert while others can,
cannot claim that humans aren’t subject to GIT
2 books published by Hubert Dreyfus (and when?)
What Computers Cant Do (1972) and What Computers Still Cant Do (1992)
Dreyfus advocated for which of Turing’s arguements
The Argument from Informality
Argument from Informality
the human mind is not constrained by rigid, formal rules like a machine
GOFAI
good old fashioned AI - all intelligent behaviour can be modelled by a system that reasons logically from a set of facts and rules
DENDRAL developeed by who, when and where
Buchanan, Feigenbaum and Lederberg at Stanford
one of the earliest expert systems
DENDRAL
2 inputs to DENDRAL
mass spectometry data and the chemical compound of a molcule
DENDRAL was powerful because what had been mapped?
theoretical knowledge needed to solve the problem
DENDRAL marked a shift from
general reaonsing (weak methods) over axioms to experts’ rules that chunk large amounts of knowledge in the domain into specific rules
DENDRAL was the first k____-i_____ system
knowledge-intensive
knowledge intensive system
expertise stemmed from a large number of special purpose rules
MYCIN (1972)
diagnose blood infections and recommended antibiotics/dosages
True/False: Did MYCIN outperform junior doctors?
yes
2 differences between DENDRAL and MYCIN
no general theoretical model from which MYCIN rules could be deduced and rules had to reflect uncertainty associated with medical knowledge
what method did MYCIN use to address uncertainty and what COULD it have used?
used certainty factors, rather than Bayesian statistics
why was MYCIN never deployed
ethical concerns and lack of electronic medical systems to integrate into
knowledge acquisition bottleneck
the difficulty of acquiring and encoding expert knowledge into an AI system
who argued about scripts?
Roger Schank
what was Roger Schanks’ emphasis when he was arguing about scripts?
less emphasis on language itself and more on representing/reasoning with knowledge required for language understanding
scripts
representations of stereotypical situations which are used to interpret stories about such situations
Schank’s Script Applier Mechanism Program
could answer based on text inference and world knowledge
scripts influenced which subdomain of applied NLP
information extracting/text mining
frame
collections of facts, procedures and default values for an object type
KL-One is an example of a
knowledge representation language
why did knowledge representation languages emerge
to support the organisation of hierarchies of frames and inference over them
knowledge representation languages were antecendants of what
OOP and description logics/ontologies/the semantic web
knowledge based systems comprise which 2 sub-systems
knowledge base + inference engine
inference engine
applies rules to a knowledge base to derive new facts/solve problems
inference engine works over _____ -rules
if-then (implication)
inference engines may use ….
forward or backward chaining
expert systems
what KBSs were called in the business world
first commercially successful KBS
R1 (later XCON)
what did R1/XCON do?
configure newly ordered computers
number of rules inR1/XCON
2500 rules
True/False: R1/XCON was 95-98% accurate
true
True/False: By 1998, most major US corps had Expert systems
True
When was the return of research funding?
1980s
how much did the UK govt put into the Alvey Project on IT
350 mil
Good old Fashioned AI (LKRR)
early AI’s emphasis on logic, knowledge representation and reasoning
3 paradigms developed in the mid 80s
Connectionism, intelligent agents, embodied/situated AI
Connectionism revived…
neural network research
key feature of Hopefield nets
recurrent - output from one unit can be fed back into itself via other units
recurrent neural nets effectively have
internal memory
recurrent neural networks can exhibit ….
dynamic temporal behaviour
hopefield nets are best suited for
unsegmented, connected handwriting tasks
RNNs use their internal memory to process…
arbitrary sequences of inputs
back propagation
sends the error at the output back thru the net to adjust the weights of connections between neurons
what would the ‘error’ be in terms of back propagation?
the difference between the network’s prediction and the actual result
who popularised back-propagation
Rumelhart and McClelland in Parallel and Distributed Processing (1986)
back propagation led to interest in…
deep learning
embodied ai
To show real intelliegence, a machine needs to have a body (perception of the real world)
which human ability does the embodied Ai approach claim is the least important?
abstract reasoning
which human ability does the embodied Ai approach claim is the most important?
commonsense reasoning
who directly attacked the physical symbol system hypothesis?
Rodney Brooks in “Elephants Don’t Play Chess
intelligent agent
an autonomous entity which observes through sensors and acts upon an environment using actuators and directs its activity towards achieveing goals
the focus of an intelligent agent is on…
integrating skills (learning, language and vision) into a single entity that can perceive and act in an uncertain, dynamic environment
internet bots are a type of…
intelligent agent
3 principal paradigms
supervised, unsupervised and reinforcement learning
supervised learning
labelled data where the
