Week 1 - lecture 1

Question 1

Why is AI not solely a psychology nor solely a computer science?

Answer 1

Understand that AI is not solely psychology or computer science but draws from both disciplines.

Recognize that AI emphasizes computation more than psychology and emphasizes perception, reasoning, and action more than computer science.

Question 2

Why AI in psychology?

Answer 2

Psychology is a big inverse problem*
AI can use forward modelling*:
we design a simple system -> see how it behaves
= the connection between AI and computational psychology

Question 3

1940s:
Warren McCulloch
Walter Pitt
2 views on artificial intelligence

Answer 3

Any computable function can be computed by a network of neurons

All logical operators* can be implemented by simple neural networks

Question 4

Warren McCulloch&Walter Pitts’three principles:

Answer 4

Basic physiology

Propositional logic

Turing’s theory of computation

Question 5

1932
John Searle
- View on Artificial Intelligence

Answer 5

A collection of cells can lead to thought/action/consciousness

Consciousness requires actual physical-chemical properties of human brains. This led to the idea that only brains can cause minds!

Question 6

Chinese room argument
- Explain

Answer 6

The Chinese Room Argument is the idea that a computer, following a set of instructions for manipulating symbols, can’t truly understand and possess consciousness.

In the scenario, a person who doesn’t understand Chinese is in a room with a set of instructions for manipulating Chinese symbols. Despite producing coherent Chinese responses, the person inside the room does not comprehend the meaning, analogous to how a computer might process symbols without genuine understanding. The argument aims to question the sufficiency of symbol manipulation for genuine understanding or consciousness.

Question 7

1950:
Alan Turing
Sentient or non-sentient AI?

Answer 7

Non sentient AI

Question 8

1950:
Alan Turing
- The imitation game
- When do we call a machine intelligent?

Answer 8

A machine is intelligent if we cannot distinguish it from a human in conversation

Question 9

1956:
Dartmouth conferences

Answer 9

Pioneers in the field of:

Computer science

Mathematics

Cognitive science

The pioneers got together for a month-long conference and coined the term artificial intelligence

Question 10

50s & 60s:
- Development in AI (2)

Answer 10

AI developed to be able to:

Play chess
Formal theorem proving

Question 11

1965:
Weizenbaum
- ELIZA machine; How does this machine work?

Answer 11

ELIZA looks for keywords in its input (father, mother, angry, happy, sad, etc.)

ELIZA uses a database of rules

ELIZA constructs new sentences, using the key words:

‘I hate my father’

‘Why do you hate your father?’

If there are no key words present in the input:

‘I see, please go on’

Question 12

1972:
Kenneth Colby
Modified Turing Test: PARRY
- How does this machine work?

Answer 12

PARRY simulated a patient with paranoid schizophrenia

PARRY was often inconsistent or meaningless sentences, but therefore realistic

Question 13

60s-70s:
AI winter

Answer 13

Overconfidence in AI systems

Led to the AI winter*

Many unanswered questions: how do we deal with perception, robotics, learning and pattern recognition

Symbolic AI does not suffice

Question 14

After the 70s:
Rumelhart
McClelland
- Revival of connectionism; development of which three kinds of connectionism?

Answer 14

Connectionism was revived:

Parallel distributed processing
Connectionism
Artificial neural networks

Question 15

1981:
McClelland
- What model did he build?

Answer 15

He build a model of human memory:

Human memory is content-adressable*

-> memory is not stored in neurons, but in the connection between neurons (Excitatory and inhibitory connections)

Question 16

1997
- Development in AI

Answer 16

First time a computer beat a grandmaster at chess in a tournament (IBM’s deep blue)

Question 17

2005:
- Development in AI

Answer 17

Last time a human beat a top chess computer under tournament conditions

Question 18

2016
- Development in AI

Answer 18

Google’s DeepMind’s AlphaGo defeated the world’s number one player in GO (An extremely difficult game)

Question 19

2020:
- Developments of AI (2)

Answer 19

Models capable of simulating general-purpose language understanding.

Large Language Models (LLM’s)

Question 20

Symbolic AI* (GOFAI)
- Definition

Answer 20

Symbolic AI (GOFAI) does not concern itself with neurophysiology

Human thinking is a kind of symbol manipulation:
IF (A>B) AND (B>C) THEN (A>C)
Intelligence = thought of as symbols and the relations between them

Conclusion:
intelligent behavior through manipulation of symbols.

Question 21

Symbolic AI* (GOFAI)
- Cons (3)

Answer 21

Why do we need it?
-> seems unnecessary for many behaviors

It is unclear how processes like pattern recognition would work in a purely symbolic way

Representations dealing with noisy input are needed

Question 22

Connectionist AI
- Definition; biologically inspired

Answer 22

Connectionism is based on the structure of the human brain.

Neurons receive input through dendrites

Neurons send output through axon

It is highly connected (1000s of synapses on axon and dendrites)

-> Computation is massively parallel

How do these neural networks compute?

Question 23

Principles of the formation of memories: (4)

Answer 23

Mental States as Vectors*, N-Dimensional Vectors

Neural Network Units*

Activation Values*

Memory and Connection Strength

Question 24

Connectionist AI
- Conclusion of definition

Answer 24

representations in the brain are distributed, processing massively parallel.

Question 25

Connectionist AI
- Pro: Lesion tolerant

Answer 25

Lesioned or damaged networks can still process informationC

Question 26

Connectionist AI
- Pro: Capable of generalization

Answer 26

ANNs are capable of learning, and are able to generalize rules to novel input.

Question 27

Connectionist AI
-Other Pro’s (4)

Answer 27

It can solve complex, non-lineair or chaotic classification problems

There is no a priori assumption about problem space or statistical distribution.

Artificial neural networks can compute any computable function.

Pattern recognition

Question 28

Large Language Models
- Definition

Answer 28

Known as a form of generative AI.

Question 29

Large Language Models
- Con’s (4)

Answer 29

False information presented by the model as fact

Compression artifacts

Also caused by the reward model of ChatGPT.

Human trainers of ChatGPT prefer longer answers over factual content.