Students should understand, by the end of the lesson, that perception is more than sensing. 

With the facial expression example, they should understand that they are seeing facial features but perceiving feelings. 

Just like how humans use “sensors” (eyes, nose, ears) to perceive, AI also perceives the world using sensors. One example is self-driving cars. Sensors like radars (a device that uses bouncing radio waves to map the world) and LiDARs (a device that uses echoing lasers to map the world), in combination with a series of cameras, are the car’s eyes.

Perception isn’t just about seeing though, being able to understand what someone is saying is also perceiving! Voice assistants like Alexa and Siri use a microphone to hear us. The microphone converts the sound waves to numbers, the numbers to phonemes, the phonemes to strings of letters and words, and the words to sentences.

Review syllabus. Mention that although not required, students can purchase the physical copies of the AI+ME series of books. 

Ask students to prepare a piece of paper and a pen or pencil. 

In this lesson, we teach two techniques of representation and reasoning — decision trees and search trees. 

In computer science terms, representations are data structures and reasoning is performed by algorithms. Data structures are different ways your data can be stored. Imagine if you had a bunch of numbers (or words, any kind of data, really) you could store them in a list or you could store them in a tree-like structure. A tree is useful if your numbers are sums of each other so a tree can represent a hierarchical structure. 

Students don’t need to know the details about data structures but by following along the “guess the animal” game, they will walk through the reasoning process step by step to guess the animal. For example, every new question is a non-terminal node which leads to other nodes, until we get to a terminal node, which gets us the answer. 

When the answer is a new animal that AI doesn’t know, AI asks questions to figure out where in the tree (AKA the representation) the new animal fits.

Page 15: After the “guess the animal” game, ask students to explain this procedure in their own words. This will help them to think deeply about the reasoning process. 

Page 17: How would you add a zebra? Ask students to draw out the tree on a piece of paper to add a zebra, then share answers with the class.

Page 21: Ask students to complete the tic tac toe game on a piece of paper and decide which strategy is the best. 

Answer: Strategy 1 is the best move. X wins in either scenario with strategy 1. 

Watch this video about AlphaZero, Google’s game-playing computer program.

Point out that AlphaZero must have a representation of the chess board then calculate every possible move and many moves after that, for every decision it makes.

Also important to know that although AlphaZero seems impressive, the games of chess, Go, and shogi are limited to a game board with defined rules. The real world is much more complex than that, many things humans do don’t have defined rules. Can you imagine AlphaZero navigate a business negotiation?

Start the class by asking students to reflect on the last time they learned something, and how they learned it. Was it by experience, by example, by observation, by plainly by being told? Certain things like playing tennis are learned primarily by experience. You can read a book on how to play tennis, but that wouldn’t help much if you didn’t practice! 

Watch a video of Google DeepMind teaches itself to walk.

Transition to the topic of today: how AI learns. Of course, to understand how AI learns, it’s important to note the differences between human learning and machine learning. 

Machines learn by finding patterns in data or optimizing behavior based on trial and error. Humans can also learn in this way, completing a task by following a series of steps. A good example of this would be solving a long division problem.

But there are a number of ways humans can learn that machines cannot, including experimentation and observation. Human learning is general and flexible (we can apply the same knowledge to different situations) whereas machine learning is achieved by specialized algorithms made to solve specific problems. 

Because of this, machine learning can get very good, sometimes even better than human performance on some tasks, for example predicting the best time to buy flight tickets. But there are only a limited number of these tasks, and even then algorithms are prone to attacks. 

Another common myth is that machine learning works so well because it works like the human brain. It is true that neural networks take inspiration from the inner workings of the brain, but the inspiration is rather abstract. Plus, we still don’t yet understand how the brain works, but we know enough to say that modern machine learning algorithms work nothing like the brain. 

In this book, the three major types of machine learning are discussed: supervised, unsupervised, and reinforcement learning. 

Supervised learning: involves a “supervisor” or a teacher, like the name suggests. Supervised learning learns with correct answers (called outputs), and trains its algorithms to get as close to the correct answer as possible. For example, remember how you learned how to divide? Your teacher showed you step by step instructions, and you practiced with problems and compared your answers with the correct solution. The more you practice, the more likely you are to nail a division problem on the test on the first try. 

In the book, the example is training an algorithm to recognize the dog Cream. Notice that all dog pictures that the AI learns on have labels. Here, because we only care about if the dog is Cream or not, the labels are “Cream” or “Not Cream”. If we give the computer enough of these images (Cream lying down, Cream running, and other poses and other dogs), the computer would eventually be able to recognize a new picture of Cream that it hadn’t seen before.

In the “summer flowers v. fall flowers” example in the book, students are stepping into the mindset of the AI agent and experiencing the pattern finding process firsthand. 

Give students some time to figure out the flowers problem before moving on to the answer.

Notice in the flowers example, we are still giving the correct label for each flower. This means we are using supervised learning. 

Alternatively, if we give the computers only the flowers with no labels, AI can still learn, but this time this would be called unsupervised learning. One technique is to group similar images together, like the orange flowers in one group and the pink ones in another. This technique is called clustering. 

The last type of machine learning is reinforcement learning. If supervised learning is learning by example, reinforcement learning is learning by experience. answer. In reinforcement

learning, the algorithm is only provided with a reinforcement signal, that indicates how well things are going. It is not told what it should do differently to make things go better; it has to figure that out for itself. 

The DeepMind video is using reinforcement learning. 

Close out class by discussing some everyday applications of machine learning. For example, Netflix and YouTube recommendations, Alexa getting accustomed to your voice…

The goal of this lesson is to have students understand why it is hard for computers to understand us, the processes involved in human-AI interaction (speech recognition and speech generation). They should also know that the Turing Test is a test for intelligence by asking machines questions. 

Start the class with this video of people testing their accents on voice assistants. 

Before the video, ask students to predict which voice assistant (Google Assistant, Alexa, Siri) they think will do the best. 

After the video, ask students what makes speech recognition so hard. Can you think of anything else other than accents? Cultural context, talking speed, and background noises can all make someone’s speech harder to decipher. For example, if a voice assistant is trained only on American English, do you think it would do as well for British accents? 

Note that voice recognition is different from speech recognition. Voice recognition refers to AI being able to understand an individual’s voice, i.e. being able to point out that it is Tim who’s speaking, not Sally. Speech recognition doesn’t pay attention to which person is speaking. 

The Turing Test

The last part of the book talks about the Turing Test, the holy grail of artificial intelligence.  The Turing Test is a test for intelligence in which you try to tell the difference between a human and a computer by asking questions. The computer passes the test if a human can’t tell if they are speaking to a computer or a real human. 

Ask students what questions they would use to test for a machine’s intelligence? How would they test if a machine can think? 

Page 5: 

Self-driving car makers: positively impacted. 

Non-self-driving car maker: negatively impacted. They might lose business as people shift to self-driving cars. 

People who cannot drive: positively impacted. They will now have more options to get around and won’t be confined to the availability of Uber/Lyft drivers.

Truck drivers: negatively impacted. Their jobs may be replaced by self-driving cars but full automation of truck driving is still far into the future. Even then, there are things that truck drivers do that are still hard to be replaced, such as loading and unloading things and customer service. 

Page 7:

Why is AI needed?

For any problem we need to consider if AI is even needed in the first place. Just because we can use AI, it doesn’t mean it is the best solution for some cases.

Who may be hurt by this change? Who may be helped by this change?

It is critical to consider how AI impacts everyone before making a decision. Since AI is powering the next industrial revolution, we need to make sure people are not left behind, especially those who are already disadvantaged. 

Next, select a few of the AI impacts in different industries to speak about. You may not have time for every industry. 

AI & Healthcare

AI can help us discover new drug treatments faster and cheaper. It costs a lot to develop a treatment for a disease because the majority of that effectively goes down the drain, because it includes money spent on the nine out of ten candidate therapies that fail before regulatory approval. 

With AI, researchers can discover patterns between biological entities such as genes, symptoms, diseases, proteins, tissues, species and candidate drugs. AI can then put all this data in context and surface the most important information for scientists.

AI & Education

Personalized AI tutors can track your learning process and predict what you need to brush up on. 

AI & Farming

Video link: https://www.youtube.com/watch?v=OhswzqyVuLw

Watch this video and discuss how AI is used in this apple picking robot. Students may say computer vision is used to gauge ripeness, path planning is used to drive up and down the field, etc. 

AI & Entertainment

AI can recommend movies and music to us on sites such as YouTube and Spotify based on what you have previously enjoyed.

AI & Social Media

AI can help select interesting content to show in social media updates. It can recognize and tag friends in your photos.

AI & Business

Chatbots are used on many websites now to answer questions. It can analyze what questions others have asked and find the best answer for you. Have you chatted with one and did you think it was helpful? 

AI & Space Exploration

We can send robots to explore space without having to worry so much about their safety. 

Sending a robot to space is also much cheaper than sending a human. Robots don’t need to eat or sleep or go to the bathroom. They can survive in space for many years and can be left out there—no need for a return trip!

Plus, robots can do lots of things that humans can’t. Some can withstand harsh conditions, like extreme temperatures or high levels of radiation. Robots can also be built to do things that would be too risky or impossible for astronauts.

Data Bias 

To illustrate how bias occurs, ask students to do the following exercise: picture a shoe in your mind. Doesn’t matter what kind of shoe, just the first that comes to your mind. 

Play the following video: https://www.youtube.com/watch?v=59bMh59JQDo

Bias can occur in a number of ways in machine learning. Consider the three stages in Teachable Machine: input, training, output / testing. Bias can seep in if the input data used to train the model is not representative of reality. Bias can also occur because of poor training algorithms, although that is relatively rare. Lastly, even if the input data is reflective of reality, we know that human bias permeates many facets of our lives, so machine learning is at risk of amplifying our existing human biases. 

Research has shown that the machines we build reflect how we see the world, whether consciously or not. For artificial intelligence that reads text, that might mean associating the word “doctor” with men more than women.