Lec 3-4

Question 1

When were the Golden Years of AI? And why is it called that way?

Answer 1

1955-1974. There was lots of funding, more promises and optimism.
Symbolic AI was the prevailing method.

The end goal was Artificial General Intelligence (building human like AI using PSS)

Question 2

What a PSS?

Answer 2

Physical implementation of system that contains symbols
Expressions consisting of symbols and their relations
Processes/rules that manipulate these expressions and are able to create new ones.

— instantiation of a symbolic AI agent

Question 3

What is the PPSH?

Answer 3

PSS has necessary and sufficient means for general intelligent action.

Human brains are also PSS and this would mean modern computers can exhibit general intelligent action.

Question 4

A logical agent being an “automated theorem prover”, how do we automate the answering of a question?

We have a KB and query formula and we want to know whether the KB entails the query formula

Answer 4

If the inference rules and complete and sound, we would find either the formula or the negation. If not, the algorithm will never terminate. - FOL is only semi-decidable

Refutation-based methods use proof by contradiction

Question 5

What are refutation-based proofs?

Answer 5

Proof by contradiction. Add the negation to the KB and show it is invalid .

Question 6

What is satisfiability?

Answer 6

We say that a formula is satisfiable if there is a model such that the model semantically entails the formula. (A given interpretation)

Question 7

What is validity?

Answer 7

The formula is valid if every model semantically entails the formula. Therefore, if there is one model that does not satisfy it, it is not valid.

Question 8

What is a clause?

Answer 8

A formula of the form L1 V .. V Ln where Li is a literal . No other logical connectives. All variables are universally quantified.

An atom is a predicate, a literal is an atom or the negation. A clause is a disjunction of literals.

Question 9

What is Conjunctive Normal Form?

Answer 9

Conjunction of clauses, a list of clauses

Question 10

What is skolemization?

Answer 10

Process of getting rid of existential quantifiers:

Replace every existentially quantified variable X with a function.
- Preserves satisfiability but not validity!

Question 11

What is resolution?

Answer 11

Resolution involves taking two clauses with complementary literals and producing a new clause. Complementary: Input has disjunction of S other has negation of S.

New clause is the resolvent of the inputs

Question 12

When would we say a list of clauses is unsatisfiable?

Answer 12

When the resolvent is the empty clause

Question 13

What are the things a term can represent in a formula?

Answer 13

A constant, variable or function

Question 14

What is a unifier?

Answer 14

A substitution of A1 and A2 if A1[0] = A1[0] (syntactically)

Question 15

What is a Most General Unifier (MGU)

Answer 15

A unifier ) is called a MGU if for any other unifier x of A1 and A2, there is a list of substitutions such that when substituting it comes down to the A1[x] = A1[0]sub = A2[0]sub = A2[x]

Question 16

Resolution Theorem Proving

Answer 16

To prove that a formula is valid:
1. Take the negation
2. Put this in clausule form and collect clauses in C
3. Find a resolution derivation of {} using unification where needed
- might never terminate
4. If the negation is unsatisfiable, the formula is valid

Question 17

What is a horn clause? A definite or goal clause?

Answer 17

At most one positive literal. (With exactly one it would be a definite clause, without it would be a goal

Question 18

What is SLD resolution?

Answer 18

It’s resolution applied to a definite clause and a goal clause. The result is always a goal clause.

Question 19

What was the focus of McCulloch & Pitts?

Answer 19

What logical functions could be encoded in an ANN. - Mostly theoretical

Learning an ANN was not yet considered.

Question 20

What is Hebbian Learning?

Answer 20

Associative learning - like reinforcement learning. Cells that fire together wire together.

Question 21

What was Minky’s SNARC?

Answer 21

Stochastic Neural Analog Reinforcement Calculator.

First ever ANN, learned parameters through Hebbian mechanisms.

Question 22

What are perceptron neurons?

Answer 22

Use numerical input and weights, bias term instead of threshold value. Activated by x0 with constant value of 1

Question 23

What is the Perceptron Learning Task?

Answer 23

Given a dataset, we want to train a perceptron so that it learns the association in the dataset. Find weights to make this possible (if possible)

Question 24

What does it mean for something to be linearly separable?

Answer 24

There is a separating hyperplane such that A and B lie on different sides. Encoded in a neuron with weights w.

Question 25

Finite sets are linearly separable if

Answer 25

Their closed convex hulls are disjoint

Question 26

What is Rosenblatts Universal Representation Theorem?

Answer 26

For any function F: {0,1}^n -> {0,1} there is an MLP with a single hidden layer, whose input-output relation is F. (For binary maps)

Question 27

What is pattern matching?

Answer 27

For any x, a corner of an n-dimensional hypercube, there exists a hyperplane that cuts off this corner. Furthermore, there exists a neuron that activates only x->

Question 28

Summary of MLP’s

Answer 28

Networks with multiple layers. A layer consists of neurons that feed their outputs into neurons in the next layer. A hidden layer is not directly connected to network input/oputput.

They can encode/approximate arbitrary functions. Results tend to be theoretical; show existence but no method to find them.

Question 29

What is considered the start of the first AI winter?

Answer 29

1969 Minsky published the book perceptrons about their capabilities:

• Single neurons can’t represent XOR (nor anything else not linearly separable)
• feedforward networks can’t count (you need recurrent connections)

Question 30

What is the Lighthill Report?

Answer 30

1973: Critical look at promises and accomplishments of AI in preceding decades. Widely cited as reason for the withdrawal of funding.

‘In no part of the field have the discoveries made so far produced the major impact that was then promised’

Question 31

What period is considered the first AI winter?

Answer 31

1974-1980