Exercise 1: Is this AI or not?
Which of the following are AI and which are not? Choose yes, no, or “kind of” was kind of means that it both can be or can’t be, depending on the viewpoint.
A spreadsheet that calculates sums and other pre-defined functions on given data
Predicting the stock market by fitting a curve to past data about stock prices
Ans: Yes, No, Kind of
A GPS navigation system for finding the fastest route
Ans: Yes, No, Kind of
A music recommendation system such as Spotify that suggests music based on the users’ listening behavior
Big data storage solutions that can store huge amounts of data (such as images or video) and stream them to many users at the same time
Photo editing features such as brightness and contrast in applications such as Photoshop
Ans: No, Kind of
Style transfer filters in applications such as Prisma that take a photo and transform it into different art styles (impressionist, cubist, …)
Exercise 2: Taxonomy of AI
Your task: Construct a taxonomy in the Euler diagram example given below shows the relationships between the following things: AI, machine learning, computer science, data science, and deep learning.
Where would you put AI? – Section B
Where would you put machine learning? – Section C
Where would you put computer science? – Section A
Where would you put data science? – Section E
Where would you put deep learning? – Section D
Exercise 3: Examples of tasks
Consider the following example tasks. Try to determine which AI-related fields are involved in them. Select all that apply. (Hint: machine learning involves almost always some kind of statistics).
Autonomous car – Statistics, Robotics, Machine Learning
Steering a rocket into orbit – Robotics
Online ad optimization – Statistics, Machine Learning
Customer service chatbot – Machine Learning
Summarizing gallup results – Statistics
Exercise 4: Definitions, definitions
Which definition of AI do you like best? How would you define AI?
Let’s first scrutinize the following definitions that have been proposed earlier:
- “cool things that computers can’t do”
- machines imitating intelligent human behavior
- autonomous and adaptive systems
There is no right or wrong answer, but here’s what we think:
“Cool things that computers can’t do”
The good: this adapts to include new problems in the future, captures a wide range of AI such as computer vision, natural language processing.
The bad: it rules out any “solved” problems, very hard to say what counts as “cool”.
“Machines imitating intelligent human behavior”
The good: the same as in the previous. Also, imitate is a good word since it doesn’t require that the AI solutions should “be” intelligent (whatever it means) and it’s instead enough to act intelligently.
The bad: the definition is almost self-referential in that it immediately leads to the question of what is ‘intelligent’, also this one is too narrow in the sense that it only includes human-like intelligent behavior and excludes other forms of intelligence such as so-called swarm intelligence (intelligence exhibited by for example ant colonies).
“Autonomous and adaptive systems”
The good: it highlights two main characteristics of AI, capture things like robots, self-driving cars, and so on, also nicely fits machine learning-based AI methods that adapt to the training data.
The bad: once again, these lead to further questions and the definition of ‘autonomous’ in particular isn’t very clear (is a vacuum cleaner bot autonomous? How about a spam filter?). Furthermore, not all AI systems need to be autonomous and we can in fact often achieve much more by combining human and machine intelligence.
Exercise 5: A smaller rowboat
Using the diagram with the possible states below as a starting point, draw the possible transitions in it(it is MUCH easier to do this with a pencil and paper than without).
Having drawn the state transition diagram, find the shortest path from NNNN to FFFF, and calculate the number of transitions on it.
Exercise 6: The Towers of Hanoi
What state should be in box 1? – State E
What state should be in box 2? – State B
What state should be in box 3? – State F
What state should be in box 4? – State D
What state should be in box 5? – State C
What state should be in box 6? – State A
Exercise 7: Why so pessimistic, Max?
Look at the game tree starting from the below board position. Using a pencil and paper, fill in the values of the bottom-level nodes where the game is over. Note that this time some of the games end in a draw, which means that the values of the node is 0 (instead of –1 or 1).
Exercise 8: Probabilistic forecasts
The weather forecast says it’s going to rain with a 90% probability tomorrow but the day turns out to be all sun and no rain. – Cannot be Concluded
The weather forecast says it’s going to rain with a 0% probability tomorrow but the day turns out to be rainy. – Wrong
Suppose you monitor a weather forecaster for a long time. You only consider the days for which the forecast gives 80% chance of rain. You find that in the long run, on average it rains on three out of every five days. – Wrong
In the United States presidential election 2016, a well-known political forecast blog, Five-Thirty-Eight, gave Clinton a 71.4% chance of winning (vs Trump’s 28.6%). However, contrary to the prediction, Donald Trump was elected the 45th president of the United States – Cannot be Concluded
Exercise 9: Odds
The odds for getting three of a kind in poker are about 1:46. – 1/47
The odds for rain in Helsinki are 206:159. – 206/365
The odds for rain in San Diego are 23:342. – 23/365
The odds for getting three of a kind in poker are about 1:46. – 2.1
The odds for rain in Helsinki are 206:159. – 56.4
The odds for rain in San Diego are 23:342. – 6.3
Exercise 10: Bayes rule (part 1 of 2)
Apply the Bayes rule to calculate the posterior odds for rain having observed clouds in the morning in Helsinki.
As we calculated above, the prior odds for rain is 206:159 and the likelihood ratio for observing clouds is 9
Give your result in the form of odds, xx: yy, where xx and yy are numbered. (Note that xx and yy do not mean that the numbers should have two digits each.) Remember that when multiplying odds, you should only multiply the numerator (the xx part). For example, if you multiply the odds 5:3 by 5, the result is 25:3. Give the answer without simplifying the expression even if both sides have a common factor.
Exercise 11: Bayes rule (part 2 of 2)
Consider the above breast cancer scenario. An average woman takes the mammograph test and gets a positive test result suggesting breast cancer. What do you think are the odds that she has breast cancer given the observation that the test is positive?
Exercise 12: One-word spam filter
Let’s start with a message that only has one word in it: “million”.
Your task: Calculate the posterior odds for spam given this word using the table above, starting from prior odds 1:1. Keep in mind that the odds is not the same as the probability, which we would usually express as a percentage.
Exercise 13: Full spam filter
Your task: Express the result as posterior odds without any rounding of the result. You may take a look at the solution of the previous exercise for help.
Exercise 14: Customers who bought similar products
Who is the user most similar to Travis? – Ville
What is the predicted purchase for Travis? – Sunscreen
Exercise 15: Filter bubbles
As discussed above, recommending news of social media content that a user is likely to click or like, may lead to filter bubbles where the users only see content that is in line with their own values and views.
- Do you think that filter bubbles are harmful? After all, they are created by recommending content that the user likes. What negative consequences, if any, may be associated with filter bubbles? Feel free to look for more information from other sources.
- Think of ways to avoid filter bubbles while still being able to recommend content to suit personal preferences. Come up with at least one suggestion. You can look for ideas from other sources, but we’d like to hear your own ideas too!
Filters as such are useful. They recommend content such as music that we like. The bubble phenomenon, where users get a biased view of facts and opinions, on the other hand, is clearly harmful. There are no “alternative facts” – the alternative of a fact is not a fact – but the information is always presented from some point of view. If we are not exposed to more than one point of view, we can easily end up holding a biased worldview.
Getting completely rid of filter bubbles is probably not a good idea. After all, we have always liked different things and been interested in different things. In the days of print media, the newspaper that we’d read was our filter bubble, which made sure that the bubble didn’t get too small.
We believe that the first step to avoiding the harmful effects of filter bubbles is to recognize when we are inside one. To this end, it is helpful if the applications we use clearly indicate that the recommendations that we see don’t represent a balanced overall view of the content. This can be achieved by letting the user explore the content also through other views than a list of recommendations. In Spotify, you can choose music based on its genre, and Netflix and HBO provide recommendations in various different categories.
Exercise 16: Linear regression
Calculate the life expectancies for the following example cases:
A – 81 , B – 73 , C – 84
Exercise 17: Life expectancy and education (part 1 of 2)
Let’s study the link between the total number of years spent in school (including everything between preschool and university) and life expectancy. Here is data from three different countries displayed in a figure represented by dots:
Answer: It is probably less than 90.
Exercise 18: Life expectancy and education (part 2 of 2)
In the previous exercise, we only had data from three countries. The full data set consists of data from 14 different countries, presented here in a graph:
Answer: Probably between 50 and 90 years
Exercise 19: Logistic regression
If you wanted to have an 80% chance of passing a university exam, based on the above figure, how many hours should you approximately study for?
Answer: 10-11 hours
Exercise 20. Elements of a neural network
Label the different components of a neuron into the diagram below. Hint: The input of the neuron comes from the left and the output goes to the right.
Synapse (connection) – D
Dendrite (input) – B
Cell body – A
Axon (output) – C
Exercise 21: Weights and inputs
In this exercise, consider the following expression that has both weights and inputs: 10.0 + 5.4 × 8 + (-10.2) × 5 + (-0.1) × 22 + 101.4 × (-5) + 0.0 × 2 + 12.0 × (-3) = -543.0
What is the intercept term in the expression? – B
What are the inputs? – A
Which of the inputs needs to be changed the least to increase the output by a certain amount? – D
What happens when the fifth input is incremented by one? – A
Exercise 22: Activations and outputs
Which of the activations described above gives:
the highest output for an input of 5? – Identity
the highest output for an input of 5? – Identity
the lowest output for an input of -5? – Identity
the highest output for an input of -2.5? – Sigmoid
Exercise 23: What is the perception of AI?
What’s the general impression you get about AI from the image search results? Is this an accurate representation of AI? Why or why not?
The Google image search with the query “AI” brought us almost exclusively brains made of circuits. After scrolling down a bit, we also got some shining white humanoid robots, often in a very pensive state. The color blue is dominant. Based on this, the impression would be that AI is about tinkering with wires and circuits, trying to build an electronic brain in some modern Frankenstein spirit. Not much color, not much fun.
We don’t think that this is an accurate representation of AI at all! AI is about solving practical, human problems, in our everyday life: better music, more interesting and important news, making new friends. Even the research side of AI, which is not what most of the Google hits are about, almost never involves working with hardware such as circuits and wires. It is mostly simply about applying sound scientific principles to understand how we can push the limits of our AI methods. It often takes a lot of talking to other researchers, writing on scrap paper, and programming. Not really much different from any other research.
Exercise 24: Implications of AI
What kind of articles (in newspapers and magazines or other popular science outlets such as blogs, …) are being written about AI – and do you think they are realistic? Do an online search about AI related to one of your interests. Choose one of the articles and analyze it.
- Mention the title of the article along with its author and where it was published (as a URL if applicable) in your answer.
- Explain the central idea in the article in your own words using about a paragraph of text (multiple sentences.)
- Based on your understanding, how accurate are the AI-related statements in the article? Explain your answer. Are the implications (if any) realistic? Explain why or why not.
Many of the articles that we studied were about the great promise of AI in different areas such as health-care, finance, customer service, transportation… you name it. A pattern that seems to repeat is that Google, IBM, Microsoft, or some of the other big players in the field have demonstrated a prototype product and the reporter is amazed by it. This tends to be combined with an estimate of the US or global market of the industry in question, which easily amounts to billions of euros.
The articles very rarely report anything about the actual techniques underlying the solutions, which is quite understandable since many readers wouldn’t be able to digest any technical details. (You would!)
A few of the articles we reviewed contain statements about AI “reading millions of pages” and “comprehending them”, but to be honest, we were actually expecting worse based on our Facebook feed. Perhaps the social media recommendations we get (based on our clicks! makes you wonder…) are of lower quality than what Google search can provide?
Exercise 25: AI in your life
How do you see AI affecting you in the future, both at work and in everyday life? Include both the positive and possible negative implications.
We genuinely look forward to what tomorrow has in store for us. At work, new assisting technologies emerge and existing ones mature to the point of being less annoying than useful. We’ll be able to complete our work more efficiently when interacting with machines takes less effort, and we can spend more time interacting with our colleagues and our loved ones.
In our everyday life, we are curious to see AI applications in entertainment such as movies and games. The ways in which we (everyone) use social media and access information online need to change, with more respect for privacy and truthfulness. An end needs to be put to the post-truth era, which is in part a consequence of filter bubbles created by AI algorithms. In this respect, we hope that the balance will tip more towards the good uses of AI and away from the bad ones. Personally, we will do our very best to contribute to this process.
Above all, we will be excited to hear back from you and learn about what we can achieve together by investing time and effort in open AI education, learning about AI, and using our improved understanding to do wonderful things.