Module 1: AI and Risk - Introduction and Overview

Question 1

Who developed the first mechanical AI machine and how was it called?

Answer 1

Charles Babbage -
Difference engine

Question 2

What is the “Turing Test”?

Answer 2

A computer proves itself to be intelligent if it can generate textual conversation of a quality indistinguishable from that of an intelligent human

Question 3

what is the difference between specific and general AI?

Answer 3

General AI - mimicking human abilities to reason
Specific AI - reasoning capabilities confined to a domain with boundaries

Question 4

Who coined the term “artificial intelligence” and when?

Answer 4

John McCarthy in 1956

Question 5

What does GOFAI stand for?

Answer 5

Good old fashioned AI

Question 6

What are the three reasons that AI development in specific domains (e.g. chess) is useful?

Answer 6

1) Games are simple compared to real life challenges
2) Explicit rules allow easier programming
3) Competitiveness allows the measurement of skill

Question 7

Which game did Arthur Samuel develop?

Answer 7

Nim - a game where two players successively remove coins with the winning side being the one which takes the last coin

Question 8

What is simple reinforcement learning (positive and negative)?

Answer 8

After each turn/move the machine updates the probabilities associated with the choice it can make at a given turn (e.g. Nim game). Positive reinforcement happens when a game is won, and negative vice versa.

Question 9

Why does a completely random player distort reinforcement learning?

Answer 9

The AI will win games that it should have lost and the random opponent will lose games that it should have won leading to inaccurate update of choice probabilities

Question 10

What is a recursive function?

Answer 10

A function that calls itself

Question 11

Which game is used to demonstrate the power of recursion?

Answer 11

The Tower of Hanoi

Question 12

When are Lookaheads useful?

Answer 12

Lookahead becomes essential
- form a multistep plan to achieve some goal
- a very large number of possible situations overall
- relatively constrained and predictable range of options

Question 13

What is an example of a search algorithm and how does it work?

Answer 13

Binary Search
Narrowing down list of possible outcomes by two with each iteration

Question 14

Which three Lookahead algorithms are presented?

Answer 14

1) Breadth-First Search
2) Depth-First Search
3) A* Search

Question 15

How does the Breadth-First Search algorithm work?

Answer 15

We explore all possible options step-by-step and eliminate in parallel sub-optimal options.
E.g. We have two ways to the same station but one is longer than the other. We eliminate the longer route and search only from the shorter route onwards as we know we can reach one station faster with with the shorter route

Question 16

How does the Depth-First Search algorithm work?

Answer 16

We explore options until the end and cross out optional pathways based on the results we received from previous paths

Question 17

How does the A* Search algorithm work?

Answer 17

We use an optimistic heuristic that we add to pathways to identify sub-optimal routes and to exclude them from calculations. We start with the most optimistic pathway after the heuristic was added. We then explore the next options and if another more optimistic pathway is identified (e.g. can be one that previously was not the most preferred one), the algorithm proceeds to check options for the new optimistic predicted pathway.

In a nutshell: With the heuristic added, we only explore options that have a chance to be more optimal than the current one under investigation

Question 18

What is Minimaxing?

Answer 18

In an adversarial game, each side tries to maximize payoff for itself while minimize payoff for the opponent. I.e. the players choose to play moves that decrease or keep the payoff for the opponent equal.

Question 19

How can games be evaluated where the decision-tree is too complex to reach the final end position?

Answer 19

We have to define heuristics which allow for an evaluation after a set of moves (e.g. depth of 6 moves). E.g. material balance, conquered space, etc.
Problem: how to weigh the different dimensions?

Question 20

What algorithm can we use to decide how to weigh heuristics for games that are too complex to solve?

Answer 20

We can use reinforcement learning.
Assign random weighs for one strategy and then deviate slightly in the weights (can also be randomly). Let the competing strategy play against the first strategy and update weighs based on the outcome of the games. In this way, an optimal distribution of weighs can be found.

Question 21

What is the curse of exponential complexity?

Answer 21

A too large set of possible choices leads to a combinatorial explosion that is too time-consuming to evaluate in order to find the optimal strategy

Question 22

What is the “frame problem”?

Answer 22

Keeping track of which aspects of a situation change when some action is performed, but also which aspects stay the same. This is important when using heuristics to assess a specific situation

Question 23

AlphaGo and AlphaGoZero use which kind of underlying method to learn how to play games?

Answer 23

Reinforcement Learning

Question 24

What four types of machine learning are used to train AI?

Answer 24

1) Reinforcement Learning
2) Supervised Learning
3) Unsupervised Learning
4) Semi-Supervised Learning

Question 25

Provide two examples of unsupervised learning

Answer 25

1) Cluster Analysis
2) Dimensionality Reduction

Question 26

When do we use semi-supervised learning?

Answer 26

Semi-unsupervised learning is used when only part of our data is labeled. The unlabeled data is is used to determine patterns within the explanatory variables or is fitted with “pseudo-labels” (determined by estimating a model on the labeled part of the data).

Question 27

Provide a method for unsupervised learning

Answer 27

Principle Component Analysis

Question 28

How does Principal Component Analysis work?

Answer 28

We have a dataset with observations and multiple attributes. E.g. measuring a student’s score in math, reading, reasoning, etc.
PCA then plots the data points in multi-dimensional space and calculates the averages of all attributes. The point of all average becomes the center of the coordinate system. Then, the PC1 is calculated as the best fitting line across all data points. After that, we calculate a line orthogonal to the best fitted line (PC2). Each further PC line will be orthogonal to the existing lines. At the end, we calculate how much each line from PC1 to PCx explains in variation. We then use only those PC lines which explain most of the variation to identify clusters.

Question 29

Why is the inscrutability of deep network machine learning problematic?

Answer 29

1) Any bias that exists in the labelling of the training data will be implicitly learned by the model and perpetuated, but in such a way that its presence is hidden and hard to eradicate
2) Risk is to privacy, because this deep implicit entanglement of complex information within the learned network can easily hold clues to personal characteristics that we would prefer to keep secret (e.g. predict a person’s gender based on online behavior)
3) Risk of manipulation arises due to the risk in privacy as personal information can be used to directly target a person

Question 30

Can the removal of attributes be of help to reduce a trained bias?

Answer 30

“Explicit” attributes that lead to bias are often correlated with other attributes so that their removal might not remove the bias

Question 31

What kind of risk a company facing when an algorithm perpetuates bias?

Answer 31

Reputational Risk

Question 32

What problem is associated with over-reliance on AI?

Answer 32

Undermine our autonomy as responsible individuals and lead to serious dangers as we neglect to develop or apply our own judgment and fail to realize when the system is getting things wrong

Question 33

Which problem occurs in software development when there is too much reliance on AI coding?

Answer 33

If the task involves any nuances (or combinations of factors) that are less common or entirely novel, then there is a serious risk that you will end up with code that has the double disadvantage of looking very plausible – and quite possibly works well in most cases – while actually being incorrect, thus making the errors especially hard to identify

Programmers may find themselves spending a considerable proportion of their time fixing “bugs,” so the use of generative AI to generate code doesn’t necessarily guarantee quality code or time savings and could expose a firm to financial risk if managed poorly