ºÚÁÏרÇø

AI stands for artificial intelligence. The words “Artificial Intelligence” are used a lot. As a result, it can be difficult to determine what AI is (and what it isn’t).  One definition that works well is this:

"AI is a computer doing something that we believe humans can do better."

What people deem as AI today, may not be what we describe as AI tomorrow. As such, AI is a moving goalpost. In other words, some people may describe a certain technology as AI, while others may choose not to describe the same technology as AI. Moreover, what people deemed as AI twenty years ago, may no longer be considered AI. However, there are some things that most people currently universally consider to be AI: “Computers drawing images from a description,” “Computers writing music given a prompt,” “Computers writing an essay,” and even “Computers driving a car.”

Note: This explanation was inspired by . For a deeper exploration of this topic, please see .

So far we know that AI is a moving goalpost, and AI encapsulates all those things that a computer does that we believe a human can do better.

Machine Learning (ML) describes one category (and a large one!) of artificial intelligence. In order to define ML, it is helpful to consider two possible approaches to getting computers to do human-like tasks. We will call these approaches (1) knowledge-based AI, and (2) example-based AI. For the sake of an example, let’s consider one task that humans complete fairly well: looking at X-ray images of bones and determining whether there is a fracture present (whether the bone is broken).

Consider Approach 1: Knowledge-Based AI

With this approach, computer-scientists provide all of the rules associated with what a fracture might look like. After speaking with many radiologists, orthopedic surgeons, and X-ray technicians, a computer scientist determines exactly what will warrant a fracture on an X-ray image. They then painstakingly program all the details of what to look for when given an image, and hope for the results of their program to produce accurate results.

The approach sounds like it should work. The thought goes, in order to learn how to do something well, talk to the experts, the people with the knowledge, and learn from them. This approach was taken by early computer scientists in the 1960s, 1970s and 1980s.

Unfortunately, this approach did not work very well. Every time computer scientists thought they had a system that could accomplish a task such as the bone-fracture classification problem, their system would make a mistake. They would speak to the experts about the mistake, and they would learn that there was something that the experts forgot to mention during interviews. The computer scientists would learn that there was an exception to the rule. There was something that they had not explicitly told the computer to look for, and, as a result, the computer would classify an item incorrectly. Artificial intelligence seemed to be impossible. Knowledge-based AI, the leading approach to making intelligent machines, had failed.

Then, in the late 1980s and early 1990s, computer scientists tried a different approach. They tried approach 2. They tried example-based AI.

Consider Approach 2: Example-Based AI

With this approach, computer scientists provided machines with lots and lots of data and hoped the machine could use mathematical feedback systems, similar to how neuroscientists believed our brains learned, to find a pattern.

Let us return to our X-ray example. Imagine thousands of X-ray images of bones were given to the computer along with the diagnosis that had been provided by physicians. It would look like this: “Here is picture 1. It is a fracture.” “Here is picture 2. It is a fracture.” “Here is picture 3. It is not a fracture.” And so on. No rules were given to the computers on how to accomplish a task. Instead many, many examples of the task being done correctly were given to the computer, and it was asked to determine the pattern itself.

Did the new approach work? Simply put, the results were astounding. The system worked; it worked really well. This new type of AI, example-based AI, seemed to be the ticket to success! In fact, it was so successful it was given a new name: Machine Learning. Machine Learning is example-based AI, the approach to building intelligent computer systems simply by giving the system data and letting it mathematically determine the patterns that exist.

Computer scientists like to use this word a lot in the world of machine learning. For example, you might hear something about the designers of ChatGPT, Claude, or Gemini having developed a new model that is more powerful than their last model, and wonder, “what even is a machine learning model?”  Simply put, a “model” is a system that has been created to take input values and produce an output prediction.

Consider the following prediction-system that may be created to help a realtor determine a house price, given the number of bedrooms that a house has. Here, the input value for the system is “number of bedrooms” and the output prediction is “house value.”

To build such a system, a realtor may collect some data and put it on a simple two-dimensional graph, with the horizontal axis representing the number of bedrooms, and the vertical axis showing the house values for houses that have sold in the past in a given market. The data might look like this:

From this data, a realtor could create the prediction system by trying to match the pattern seen in the data. One such system might look like this:

Now, this particular prediction-system is one you may be familiar with: it’s a line of best fit. This prediction-system is a model. In fact, it is one of the simplest machine learning models we can make. 

What makes a model like this useful, is that we can make predictions about new data, simply by using our previous data and pattern, and a new input variable. In this case, a realtor may want to sell a house in the market that is represented by this model, and by providing the number of bedrooms, they can predict the selling price of the house. This might look like this:

Notice that there was no previous data on the cost of house for the input number of bedrooms, but the model was still able to make a prediction. This is a simple example of a machine-learning model making a prediction. A more complex model could use two variables as inputs, say, number of bedrooms and number of bathrooms, to make a prediction about house price. An even more complex example could have ten input variables to make a prediction. Very complex models may have thousands or even millions of input variables to help make a prediction. Even though we can’t visualize this easily on a graph, we hopefully can understand how input variables, accompanied by a model, can help us make predictions.

AI and Machine Learning is used everywhere to help humans make decisions. Because computer scientists have helped computers learn patterns from large amounts of data, they can provide large amounts of data on any problem and build models to help make predictions. Below are some of the areas where AI helps inform societal decision making.

Disaster Relief and Resource Allocation

Humans can provide computers with drone footage of flood areas, where humans have labeled pictures of building structures as completely destroyed, partially destroyed, or undamaged. AI models can be built to be used after a storm hits, and feed unlabeled pictures of structures taken from drones. Models can easily predict which pictures represent the most damaged homes and prioritize sending more aid to these regions, even without humans asking for help. Click here to learn about how AI can assist with resource allocation during disasters.

Medical/Public Health

Humans can provide computers with information about patients along with diagnoses and create models that can predict a correct diagnosis for a patient simply given their presenting symptoms. Another model could be used to predict the way that diseases spread in a population given information about past diseases and the patterns that are learned. . 

Transportation

Humans can provide information from vehicles with sensors that record how cars successfully drive on a road given a set of conditions and create a driving model. This model can be used for a car to predict how to drive itself. . 

To better understand AI, it is helpful to understand the terminology used. Here are some key terms to get you started:

• Big Data: Complex datasets that exceed the processing capacity of traditional data management tools. Big data is characterized by volume (a lot of data), velocity (the high speed of data creation and processing), and variety (diverse data types like text, images, and video).
• Input: What goes into our model to help us predict, often called the features (ex. number of bedrooms in a house)
• Black box: In engineering, the term is used to describe any device whose inner workings are hidden from the user. In ML, this may be a particular algorithm or set of steps, whose inner workings are mysterious to the user. A black box allows us to hide unimportant details and is very important for human problem solving.
• Output: The prediction from our model, often called the label (ex. the price of the house)
• Model: The “machine” associating the input to a predicted output by finding the pattern in the data.
• Training: When a model is being built, and data is provided to help it learn the pattern, we say the model is being “trained”. 
• Inference: When a model is being used to make a prediction about data it has not yet seen (or has not been trained on), we say the model is making an “inference”.

For a more comprehensive list of industry terms, please refer to .

Note: While many of these are available to the general public, some of the links below require subscription accounts.

The Fourth Industrial Revolution and Internet of Things (IOT)
Case Study: AI in the Film Industry
Human Decision Making and Sports

Understanding an Artificial Neuron

A single artificial neuron is the building block of a neural network. Having students understand how a single neuron can take in inputs and produce outputs is the foundation for understanding how these neurons can be composed and work in concert for more sophisticated tasks.

This activity explains how inputs, weights, activation functions, and outputs can be used to make a simple prediction about whether a new song will be a viral hit. The activity ends with a “plot twist” requiring students to consider how the parameters of the neuron must be altered to account for new data about what makes a song popular.

Materials Needed: Printed out copies of the first six pages of this document. (Teacher guide is on the final pages).

Learning Objective: Understand how an artificial neuron functions to predict an output given an input.

TensorFlow Playground

Link: 
Cost: Free

Learning Objectives:

• Key Concept: Students will understand how different constructions of artificial neurons in a neural network help create better predictors. Students will see visually how neural networks are constructed and work
• Real-World Application: Neural networks are used everywhere, from vision systems in autonomous vehicles to large language models. Understanding how these systems are built and scale
• Critical Thinking: See how more complex datasets require more neurons and layers within a network to allow the neural network to find patterns. See the limitations of smaller architectures for neural networks as well.

💡 Activity Idea: TensorFlow Activity

• Time: 40-60 minutes
• Activity: The TensorFlow Activity requires students to engage with the TensorFlow Playground to better understand how neural network architectures influence a neural network’s capabilities.

Understanding Neural Network Optimization

Neural Networks continually update to make better predictions given new information from the data. Playing a “hotter, colder” game with students can help students understand the basic terms involved in neural network learning: loss function, optimizer, epochs, and learning rate.

Learning Objectives:

• Key Concept: Understand the basic terms involved in neural network learning, including loss function, optimizer, epochs, and learning rate.

• Continual Updates: Explore how neural networks continually update to make better predictions given new information from data.

• Algorithmic Analogy: Visualize how human optimization strategies mirror machine learning optimization processes.

💡 Activity Idea: Hotter-Colder Game

• Time: 5-10 minutes
• Activity: The Hotter-Colder game will require a single student to leave the classroom while an object within the room is hidden. They will perform the task of finding the object after entering the room three times, with different rules each time about the number of steps they can take, and the size of the steps they can take as they journey around the room. Other students must say “hotter” or “colder” to the searching student and the student is said to have “found” the object when they are within some teacher-specified distance of the object and can visualize it plainly. Slides for the activity can be found here, which not only detail the process, but also explain how the game process is analogous to key terminology.

Understanding Convolutional Neural Networks (CNNs)

Neural Networks continually update to make better predictions given new information from the data. Playing a “hotter, colder” game with students can help students understand the basic terms involved in neural network learning: loss function, optimizer, epochs, and learning rate.

Learning Objectives:

• Key Concept: Understand that Convolutional Neural Networks (CNNs) are specialized neural networks used chiefly in vision systems.

• Image Interpretation: Compare how images that computers see in vision systems are different from the images humans see.

• Strengths & Limitations: Evaluate the functional strengths and inherent limitations of machine vision compared to human vision.

💡 Activity Idea: Three-Part Lesson on the Basics of CNNs

• Time: 5 minutes for part 1, 10 minutes for part 2, 50-60 minutes for part 3
• Activity: This three-part educational activity guides students through the complexities of how both humans and machines interpret visual data.
  
  

  Part 1: Understanding Human Vision Systems – Explores the biological and cognitive processes of human sight, using "I-Spy" style activities to demonstrate how we identify objects in complex environments.

  
  

  Part 2: Understanding Operations – Breaks down the logic of vision, likely focusing on the step-by-step operations required to process images.

  
  

  Part 3: Taking it to the Streets – Applies these concepts to real-world scenarios, examining how vision systems function in practical, everyday settings like autonomous navigation.

Neural Networks used in machine learning work by connecting many artificial neurons together. As shown in the diagram, neurons are organized into layers with the first layer being called the “input layer”, the last layer being the “output layer”, and the layers in the middle being called “hidden layers”. The input layer is where the data is input into the neural network (say a picture of a dog), and the output layer represents the result of the input (say, a classification of the picture as “dog”). Each neuron in the hidden layers has parameters that can be altered with names like “weights” and “bias” that change whether a given neuron will produce an output signal based on a given input signal. If all of the parameters are set correctly, input data should produce the correct output result.

Consider the neural network depiction below:

Each neuron in the depiction of the neural network has been replaced by a turnable knob (in this analogy, turning a knob is akin to adjusting parameters of weights and bias of the neuron). With a given input, and the knobs adjusted in their starting positions, the output result may not match what is expected. For example, data for a picture of a dog could render a result of “cat” with the knobs in the current positions. 

However, one could imagine that for a given input (like a picture of a dog), these knobs could be adjusted until the correct output results (the result at the output layer says, “dog”). This might look like this:

Once the neural network parameters have been set by taking many inputs and adjusting the “knobs” until the correct output appears, it can be used on data that has not been seen before to predict the output based on the input. 

The process neural networks use to adjust these parameters requires feeding data forward through the network (in a process called “feed-forward”), and then calculating the error (or loss) to update/adjust the parameters of the neurons in the network by moving backwards (in a process called “back-propagation”). The network learns by minimizing this loss. Each time the network feeds forward data and backpropagates it is called an epoch. Generally, the more epochs a neural network has to learn, the better the results will be. The amount that the parameters are adjusted at each pass is called the learning rate, and higher learning rates mean that a neural network may learn faster, but not necessarily converge to the correct result. The hotter/colder activity above in the Learning in Action section of this module helps explain this phenomenon.

The analogy of the “turning knob” neural network helps us understand how the neurons are individually adjusted to tune into the correct result, but it is important to recognize that neural networks do this work of adjusting the parameters automatically using sophisticated mathematics (involving the multi-variable chain rule quantities like the gradient which can be learned about in more detail in an advanced calculus course).

Neural networks in applications today are massive in size. The number of neurons is on the order of billions or more, making them capable of handling complex tasks like image recognition or sophisticated classification tasks.

Explainability

Because of the massive size of neural networks used for tasks today, and the automated way in which they learn to correct the weights and biases to produce an output, it can be very difficult to understand why a neural network produces a result given a particular input. While the values for parameters can be determined, the practicality of understanding how billions of numbers work in concert is not just daunting, but untenable. As a result, so long as the correct results are obtained, the inner workings for how a result is obtained is often unclear.

Since it is not reasonable to directly alter a neural network parameters or even understand the precise impact of the parameters on the outputs, it is extremely important that neural networks are tested on a variety of inputs to verify the results of the outputs. Ensuring that unique cases are included in training neural networks so that they perform well for a variety of scenarios or inputs is critical. This is especially true in high-stakes situations like autonomous vehicle travel. As just one example, training a neural network for a self-driving car on country roads alone would not mean that this neural network would do well at predicting how to drive a car well in an urban environment. The power of neural networks allows them to accomplish amazing feats, but their predictive power always relies on the quality and breadth of the data used to build them.

In summary:

• If a model's inner workings are unclear: we cannot rely on logic alone to predict how it will behave in the real world.

• If a model is only trained on limited scenarios: it cannot be trusted to handle unique, high-stakes environments safely.

• If a model is thoroughly tested on diverse data: we can gain confidence in its outputs, even if we cannot map out every single internal calculation.

When we deploy complex neural networks, we must always implement rigorous, varied testing to verify the outputs of a system we cannot completely see inside.