Supervised Learning

In supervised learning, you train the AI by mapping inputs to outputs. We use labeled data to build the model. Labeled means that along with the example, we also provide the expected output. For instance, if we want to build a system that detects product defects in a factory, we can show the ML model many examples of defective and good-condition products. We can label these examples as “defective” or “good condition” for the model to learn. We can label a face recognition dataset by indicating whether an example image contains a face or not.

This technique is a key part of modern AI, with the most value created by supervised learning algorithms. Supervised learning is a common technique that accounts for about 95% of all practical applications of AI, and it has created a lot of opportunities in almost every major industry.¹

However, today, to teach a computer what a coffee mug is, we show it thousands of coffee mugs, but no child’s parents, no matter how patient and loving, ever pointed out thousands of coffee mugs to their child.³ The challenge with supervised learning is that it needs massive amounts of labeled data. Getting accurate and useful labeled data can be costly and time-consuming. Companies often employ human annotators to label training data.

A diagram of the two roles of A I models in supervised learning. On the left side, in the left column, from top to bottom are clip arts of apple, orange, and banana, which, upon training, result in an A I model. On the right hand side are a group of different fruits, which through an A I model, are now categorized according to their classifications with labels.

Unsupervised Learning

Most people would agree that if you have to teach a child to recognize cats, you don’t need to show them 15,000 images to understand the concept. In unsupervised learning, there is no labeled data. We teach machines to learn directly from unlabeled data coming from their environments. This is somewhat similar to how human beings learn. With unsupervised learning, you essentially give the algorithm lots of data and ask it to find patterns in that data. The AI learns by finding patterns in the input data on its own.

A diagram of the use of an A I model in clustering three fruits, which are apples, bananas, and oranges. From left, are the fruits that go through finding patterns, then to an A I model that clusters according to classification, without any indicated label.

Many applications of unsupervised learning use a technique called clustering. Clustering refers to organizing data by some characteristics or features. These features are then used as labels that the model generates. This technique is especially useful when you want to group things like a set of images by people, animals, landscapes, etc. or find groups of similar users for targeting ads.

On the left hand side is a pile of potatoes of different sizes and types. On the right hand side is a photograph of sorted potatoes according to their types. A horizontal arrow between the two indicates that the sorting has been completed.

Access to labeled data is often a bottleneck for many AI projects, and this is also one of the most difficult challenges facing the field. Unsupervised learning is an extremely valuable technique and represents one of the most promising directions for progress in AI. With unsupervised learning, we can imagine systems that can learn by themselves without the need for vast volumes of labeled training data.

Reinforcement Learning

Sometimes it can be difficult to specify the optimal way of doing something. In reinforcement learning, the AI learns by trial and error. In this technique, AI is rewarded for positive behavior and punished for negative behavior.

We show the AI data, and whenever it produces correct output, we reward it (sometimes in the form of scores). When it produces incorrect results, we punish it by not rewarding it or deducting rewards. Over time the AI builds a model to get the maximum reward.

Training a pet is a good analogy for understanding reinforcement learning. When the pet displays good behavior, you give it treats, and when not, you don’t give it any treats. Over time the pet learns that if they behave in a particular manner, they will get rewards, and you get the pet’s good behavior as an outcome.

A diagram of a silhouette of a standing cat, on the upper right side, which is in positive reinforcement, can play a ball, or receive a reward. While on the lower right hand side, with negative reinforcement, it will just sit, and will not receive a reward.

Reinforcement learning is used for building AI that drives autonomous vehicles like self-driving cars, drones, etc., as well as for playing games like chess, Go, and video games. The AI can play a lot of games against itself to learn on its own. Google’s DeepMind team trained their AI to play 49 different Atari video games entirely from scratch, including Pong, Freeway, and Space Invaders. It used only the screen pixels as input and the game score as a reward signal.⁴

The main challenge with reinforcement learning is that it requires a large number of practice runs and massive computing resources before the algorithm is usable in the real world. A human can learn to drive a car in 15 hours of training without crashing into anything. If you want to use the current reinforcement learning methods to train a car to drive itself, the machine will have to drive off cliffs 10,000 times before it figures out how not to do that.⁵

Deep Learning and Neural Networks

Deep learning (DL) is a subset of machine learning. In ML, learning happens by building a map of input to output. This map is the ML model represented as a mathematical function.

The mapping can be straightforward or convoluted and is also called a neural network. For example, clicking a button to switch on a light bulb is an easy problem that a simple neural network can accomplish. The input is the click, and the output is the state of the light bulb (on/off). Learning this sort of mapping is more straightforward and sometimes referred to as shallow learning. For more complex cases like predicting the price of a house based on input parameters like size, the number of rooms, location, distance from schools, etc., learning happens in multiple steps. Each step is a layer in the neural network. Many of these steps are not known and are called hidden layers. Learning with multiple layers is known as deep learning.

In ML, depth refers to the number of layers in the neural network. Networks with more than three layers are generally considered deep. While the term neural networks might sound brainlike, neural nets are not trying to imitate the brain, but they are inspired by some of its computational characteristics, at least at an abstract level.⁶

DL is uniquely suited to build AIs that often appear human-like or creative, like restoring black-and-white images, creating art, writing poetry, playing video games, or driving a car. Deep learning and neural networks is a broad field and the foundation of many popular AI techniques.

A diagram of deep learning and neural networks. From left to right are input layers like size in square feet, number of rooms, location, nearest school, crime rate, and age, which are connected through lines and circular points called hidden layers, with an output of estimated price.

Backpropagation

A large part of what machine learning teams do is called “feature engineering.” Learning is enabled by a collection of what are called “hyperparameters”—an umbrella term that refers to all the aspects of the network that need to be set up by humans to allow learning to even begin.⁷ This can include how much weight we should assign for different inputs, the number of layers in a neural network, and many other variables. This is called tuning the algorithm.

An ML engineer might build a model and assign some weights to inputs and different layers. After testing the model on some data, they might adjust these weights to get better results, and they might do this multiple times to arrive at an optimal weight.

By using the backpropagation algorithm, ML teams can start adjusting weights automatically. It’s a way of tinkering with the weights so that the network does what you want. As a neural network is trained, information propagates back through the layers of neurons that make up the network and causes a recalibration of the settings (or weights) for the individual neurons. The result is that the entire network gradually homes in on the correct answer.⁸

Backpropagation is used wherever neural networks are employed, especially for speech recognition and speech synthesis, handwriting recognition, and face recognition systems.

A diagram of backpropagation. From left to right, input layers are size in square feet, number of rooms, location, nearest school, crime rate, and age, connected through lines and points called hidden layers, with an output of estimated price. Lines of varying square footage are linked by feedback and a feed-forward that directs to the output.

Transfer Learning

If a person knows how to drive a hatchback car, they can also drive an SUV. Software engineers who know a particular language like JavaScript can transfer some of their knowledge when writing code in Python. If I give you a new task, you wouldn’t be completely terrible at it out of the box because you’d bring some knowledge from your past experiences with handling similar tasks.

Transfer learning is the ability of an AI program to transfer what it has learned about one task to help it perform a different, related task.⁹ Humans are great at doing this. For people, transfer learning is automatic. In transfer learning, instead of training your model from scratch, you can pick up some learnings from another trained model that is similar. For example, an image recognition system trained to recognize cars from images can also transfer its learnings to recognize golf carts. This technique is very valuable, and many computer vision and natural language processing (NLP) systems are built using transfer learning.

Photograph A: There are four cars on a road, each in a square with a label on the upper left: Car 0.94. Photograph B is a double-decker bus on a highway, in a square with a label on the upper left: Bus 0.86. Between the two photographs is a horizontal arrow.

Generative Adversarial Networks (GANs)

Perhaps you’ve seen examples of AI capable of generating images of people who don’t really exist, AI-generated speech that sounds human-like, or AI restoring black-and-white pictures to color. These are all examples of generative adversarial networks or GANs in action. Researchers at Nvidia used GANs to generate photorealistic images of fake celebrities from pictures of famous people on the Internet.

Generative adversarial networks were developed by Ian Goodfellow and have been called the most interesting idea of the decade by Facebook’s chief AI scientist, Yann LeCun.¹⁰ While explaining how GANs work is out of the scope for this book, at a high level, this technique is a subset of deep learning that generates new data by pitting two AI algorithms against each other, hence adversarial.

Deepfakes, where a person in a specific kind of media—like an image, video, sound, etc.—is swapped with another person, are one of the most popular (and dangerous) applications of GANs. GANs have many applications in the VFX and entertainment industries. You can use GANs for various tasks like synthesizing new images from scratch, generating videos (or deepfakes), restoring images, self-driving cars, video game AIs, and sophisticated robotics.

On the left side are collage photographs of 25 faces of random men and women. On the right side is a collage of four house landscape paintings.

Knowledge Graphs

When you search something like “Mumbai” on Google Search, you’ll notice that Google seems to understand that Mumbai is a place and will show you a rich view of places to visit, flights, hotels, etc. This rich view is enabled by a technique called knowledge graphs.

A knowledge graph, also known as a semantic network, represents a network of real-world entities—that is, objects, events, situations, or concepts—and illustrates the relationship between them.¹¹ These rich graphs of information are foundational to enabling computers to develop an understanding of relevant relationships and interactions between people, entities, and events.¹² Knowledge graphs provide a lot of economic value and are probably one of the most underrated AI techniques.

On the left side are search results for Socrates, arranged from top to bottom, which are Images, Name and Occupation, Wikipedia, General, Books, and Relationships. On the right side is a diagram of Socrates, connected with his background information and people related to him.

Accuracy

Accuracy is the fraction of predictions our model got right. It is the proportion of correct predictions out of all predictions.

Accuracy = (True positives + True negatives) ÷ All predictions*

*All predictions = True positives + True negatives + False positives + False negatives

In the case of our donut detector, that would be

Accuracy = (70 + 21) / (70 + 5 + 4 + 21) = 0.91 (or 91%)

91%! That’s not so bad. However, there are problems with determining an algorithm’s performance based on accuracy alone. A 91% accuracy may be good or terrible depending on the situation. (It’s not good if a car crashes 9 out of 100 times).¹⁴ To understand the AI’s performance better, AI teams use metrics like precision and recall that describe the breadth and depth of results that your AI provides to users and the types of errors that users see.¹⁵

Precision

Precision is the proportion of true positives correctly predicted out of all true and false positives. Precision determines how much you should trust your model when it says it’s found something. The higher the precision, the more confident you can be that any model output is correct.¹⁶ However, the tradeoff with high precision is that the model might miss out on predicting some correct results, that is, you will increase the number of false negatives by excluding possibly relevant results.

Precision = True positives ÷ (True positives + False positives)

In the case of our donut detector, that would be

Precision = 70 / (70 + 5) = 0.933 (or 93.3%)

If the donut detector was optimized for high precision, it wouldn’t recommend every single image of a donut, but it would be highly confident of every donut it recommends. Users of this system would see fewer incorrect results but would miss out on some correct predictions.

Recall

Recall is the proportion of true positives correctly predicted out of all true positives and false negatives. The higher the recall, the more confident you can be that all the relevant results are included somewhere in the output.¹⁷ The higher the recall, the higher is the overall number of predictions. However, the tradeoff with high recall is that the model might predict more incorrect results, that is, you will increase the number of false positives by including possibly irrelevant results.

Recall = True positives ÷ (True positives + False negatives)

In the case of our donut detector, that would be

Recall = 70 / (70 + 4) = 0.945 (or 94.5%)

If we optimized the donut detector for high recall, it would recommend more images of donuts, including ones that don’t contain donuts. The user of this system would see more results overall; however, more of those results might be incorrect.

Precision vs. Recall Tradeoff

In most real-world scenarios, you will not get a system that both is completely precise and has a hundred percent recall. Your team will need to make a conscious tradeoff between the precision and recall of the AI. You will need to decide if it is more important to include all results even if some are incorrect, that is, high recall, or minimize the number of wrong answers at the cost of missing out on some right ones, that is, high precision. Making this decision will depend on the context of your product and the stakes of the situation. Weighing the cost of false positives and false negatives is a critical decision that will shape your users’ experiences.¹⁸ For example, in a banking system that classifies customers as loan defaulters or not, you might be better off making fewer but highly confident predictions. It would be better to optimize this system for higher precision. On the other hand, a music streaming app occasionally recommending a song that the user doesn’t like isn’t as important as showing a large selection. You might optimize such a system for higher recall.

You will need to design your product’s experiences for these tradeoffs. Make sure to test the balance between precision and recall with your users.

On the left, is an image of several smileys, which means true positive, and a sad face, which means false positive. Inside the circle are all smiley faces. Outside the circle, there are 3 scattered sad faces and 2 smileys. While on the right side, along with the 7 smileys, are 2 sad faces inside the circle, and 4 sad faces scattered outside the circle.

Computer Vision (CV)

Computer vision (CV) is the ability of AI systems to see by enabling them to derive meaningful information from digital images, videos, and other visual inputs. AI systems can use this information to make predictions and decisions and take action. For example, Airware was a start-up that conducted an automated analysis of drone images for mining and construction industries. They used computer vision on drone imagery to analyze safety compliance, maximize fuel efficiency, calculate stockpile volumes, and recognize storm damage.¹⁹ Google Lens uses computer vision algorithms to identify what’s in a picture.

Natural Language Processing (NLP)

Natural language processing (NLP) is the ability for AI to pull insights and patterns out of written text. NLP is one of the most valuable capabilities for AI systems and is used in many applications like AI assistants, search engines, chatbots, sentiment analysis, translation, etc. For example, voice assistants like Alexa, Siri, Cortana, or Google Assistant use NLP to act on your commands. Everlaw is a cloud-based AI application that helps lawyers research documents and prepare for trials by extracting information from legal documents, highlighting which documents or their parts are important. They automate repetitive tasks for lawyers and paralegals by identifying documents relevant for trial by topic, translating documents into other languages, clustering documents by topics, and audio transcription.²⁰

Speech and Audio Processing

This is the ability for AI to convert speech to text and text to speech (TTS) and extract information from audio files. Speech processing is often used alongside NLP in various AI applications like voice assistants, call center analytics, automatic subtitle generation, etc. For example, voice assistants like Alexa, Siri, Cortana, or Google Assistant use speech processing to recognize and convert your voice into text. YouTube uses speech processing to generate automatic subtitles on videos. Spleeter is an open source software made by the music streaming service Deezer that uses speech and audio processing to automatically separate vocals and different instruments from songs.

Starting from the left, Image A is the welcome page of Spleeter, which has 4 categories: the vocal track, drum track, bass track, and other track. B is a clipart image of a man that gives verbal details, converted to written text. C is a photo of an Amazon Alexa device.

Perception, Motion Planning, and Control

Robotics is a branch of technology that deals with the design, construction, operation, and application of robots. Primarily used in robotics, this is the ability of AI systems to plan and navigate spaces through sensors and control actuators. An actuator is a component of any machine that enables movement like a robot’s limbs, steering wheel in a self-driving car, propellers on a drone, etc. Perception helps a robot figure out what’s around it, motion planning allows the robot to map the best route, and control allows it to send commands to actuators. Robots are used in many industrial applications like warehouse robots, assembly lines in factories, and delivering products using drones.

Photographs A, B, and C depict a robotic vacuum cleaner, two robotic arms, and a two-seater self-driving car, respectively.

Prediction

Prediction is the process of filling in the missing information. Prediction takes the information you have, often called “data,” and uses it to generate information you don’t have.²⁵ This is one of the most valuable and powerful capabilities of AI systems. For example, AI models are used in systems that predict flight prices or the number of orders while managing inventory. A ubiquitous example is the smartphone keyboard that uses predictive models to suggest the next word. Furthermore, predictions can be used to make an investment, detect and correct health disorders, or choose the right word in a situation.

Ranking

You can use AI to rank items, especially when it is difficult to determine a clear ranking logic. Ranking algorithms are used in search engines to decide the order of results. PageRank is an algorithm used by Google Search to rank web pages in search results. Ranking is also used along with other AI applications like product recommendation systems to decide the order of items suggested.

Classification and Categorization

This is the ability of AI to categorize entities into different sets. Categorization can be used for several applications like sorting vegetables from a pile or detecting faulty products in an assembly line or by photo apps to classify images into landscapes, selfies, etc. Clustering is a common technique used for generating a set of categories. For example, in an advertising company, you can use clustering to segment customers based on demographics, preferences, and buying behavior.

Knowledge Representation

Knowledge representation is the ability of AI systems to extract and organize information from structured and unstructured data like web pages, books, databases, real-world environments, etc. This capability enables AI systems to understand relevant relationships and interactions between people, entities, and events. A rich search engine results page (SERP) is an example of knowledge representation in action. For example, when you search for a famous person like “Alan Turing” on Google Search, the result shows you rich information about his date of birth, partner, education, alma mater, etc.

On the left is a search bar, with a search for running shoes and predictions below. Besides it, there is a page of available flight bookings for different airlines, sorted based on price. On the right hand side, there is a search result in images, texts, and related searches for Socrates. At the bottom is a page with a search bar and recommended photographs below.

Recommendation

Recommendation is the ability of AI systems to suggest different content to different users. For example, Spotify suggests what songs to listen to next, or Amazon recommends which books to buy based on previous purchases and similar buying patterns. Recommendation is a very valuable capability and is closely related to prediction. You can also use recommendations to surface content that would otherwise be impossible for users to find on their own. Many social media applications like TikTok or Instagram use recommendation algorithms to generate a continuous stream of dynamic content.

A is a page on Netflix with a list of movie or series recommendations. Below, on the left part, are recommended articles on Medium homepage, and on the right side are book recommendations on Amazon.

Pattern Recognition

This is the ability of AI systems to detect patterns and anomalies in large amounts of data. Detecting anomalies means determining specific inputs that are out of the ordinary. For example, AI systems can help radiologists detect lung cancer by looking at X-ray scans. Banks use software that detects anomalous spending patterns to detect credit card fraud. Market researchers can also use pattern recognition to discover new segments and target users.

Eight X-ray images of different analyses of human lungs. Atelectasis is circled on the upper right part, while on the center part is cardiomegaly. The bottom right is effusion. On the central left part are infiltration, mass, and nodule. In the center-right part is pneumonia, and in the upper left part is pneumothorax.

Apart from the preceding capabilities, machine learning is also used broadly to find patterns in data. Many of the examples discussed previously relate to unstructured data. The popular press frequently covers AI advancements applied on unstructured data like detecting cats from videos, generating music, writing movie scripts, etc. These examples often appear human-like and are easily understandable by most people.

However, AI is applied at least as much or more to structured data, that is, data is a standardized format for storing and providing information. Structured data also tends to be more specific to a single organization. AI on structured data is creating immense economic value to organizations. But since it is harder for popular media to understand AI advancements on structured data within companies, it is written about much less than AI advancements over unstructured data.

AI is evolving, and many new techniques will unlock newer capabilities. A large part of what AI teams do is figure out problems to solve with AI and match appropriate techniques and capabilities to the solution. Doing this well requires lots of iterations, time, and effort. Most large AI products combine multiple capabilities to provide a solution. For example, an AI voice assistant might use speech processing to listen for voice commands and convert them to text, NLP to understand and act on the command, and, finally, speech synthesis to respond to the user. Many of the preceding capabilities might be further combined to realize new AI methods. For example, you could extend the idea of machine translation from translating languages to translating images into words through automatic captioning, generating poetry, or even translating words back to images by creating AI artworks.