Book Review: Prediction Machines

I had been searching for a book that explained the various artificial intelligence related terms in a very reader-friendly way, and I think I found it: Prediction Machines by Agrawal et. al. 

Update: Joshua Gans, one of the authors of this book, kindly answered some questions I asked him. Read the Q&A.

The book starts with the authors explaining that they are not computer scientists, but economists. I believe this reason alone sets them apart: they touch upon the ugly details behind AI, but only to explain what these prediction machines do. Chapter after chapter, you get to read simple and clear text that explains the whole AI phenomenon. 

"Economists view the world differently than most people. We see everything through a framework governed by forces such as supply and demand, production and consumption, prices and costs."

The authors assert that since it is simply economical for us to delegate various tasks to machines, which have become better over time at predictions, we are seeing a wider adoption. While this is true for any automated task, predictions support decision making and thus are valuable to managers everywhere.

These predictions are not only accurate, often beating experts, but also are faster. Now, cheaper. Decision makers are attracted to AI because of this combination. 

Let’s see a chapter by chapter summary.

Chapter 3: Prediction Machine Magic

This chapter provides examples of how fraud prediction and language translation (powered by deep learning) are widely used. It also provides how the predictions have become better and surpassed human benchmarks.

“By 2017, the vast majority of the thirty-eight teams [participants in prediction an image competition] did better than the human benchmark…. Machines could identify these types of images better than people.”

Chapter 4: Why It’s Called Intelligence

Although many machine learning techniques share common principles with statistics, machine learning is different as this chapter explains. While regression tries to be accurate on the mean (i.e. “regress to the mean”), machine learning tries to improve the overall predictions.

To make a case for prediction machines be equal to artificial intelligence, the authors quote Jeff Hawkins, the author of On Intelligence.

“Prediction is not just one of the things your brain does. It is the primary function of the neocortex, and the foundation of intelligence.”

Chapter 5. Data is the New Oil

This chapter provides multiple examples how data in various forms have helped people build better prediction models. An example, a startup called Cardiogram built an iPhone app that receives the heart rate data from an Apple Watch. Then the company can “detect an irregular heart rhythm with 97 percent accuracy using its deep neural network.”

The authors also walk you through the basic principles of model building for predictions. They also explain more data is sometimes better as it allows companies to differentiate their products, hence “data is the new oil.”

“Increasing data bring disproportionate rewards in the market.”

Chapter 6. The New Division of Labor

This chapter shows how our biases generate inefficient or inaccurate solutions -- and that’s the strength of a machine which creates solutions without any bias. (unless, of course, its designer introduced some biases.) One particular example is worth reading about: machine predicts better than judges whether a defendant would commit a crime while on bail.

Machines, of course, aren’t infallible: the inherent biases in the training data as well as the missing data that’s not fed can generate highly inaccurate results. So, although machines can scale and create complicated models, human judgement is still required to fine-tune or guide the machine.

Phew!

Chapter 7. Unpacking Decisions

This chapter reiterates that machines provide better and faster predictions, however, the decision is still need to made after assessing risks and rewards. That decision making is still a human domain and most likely will remain for the near future. (The authors provide one painful example of the London’s cabbies.) 

“As prediction becomes better, faster, and cheaper, we’ll use more of it to make more decisions, so we’ll also need more human judgement and thus the value of judgement will go up.”

Phew++

Chapter 8. The Value of Judgment

This chapter explains how although the final decision-making is largely left to humans, by calculating the risk and rewards, some decisions can be left to the machines. The most common example of this is blocking a credit card transaction that a machine thinks is likely fraudulent. The costs of such a decision is customer irritation and loss of goodwill and the reward is the money saved on a fraudulent transaction.

Chapter 9. Predicting Judgement

Sadly, I felt there wasn’t any new information in this chapter. The authors reemphasized that machine can’t predict what they can’t see and humans are better at context. One great example of human ingenuity was of Abraham Wald and his questioning of the missing data on bomber planes during World War II.

Chapter 10. Taming Complexity

If the decision-making process is seen as a big flow chart of ifs and thens, then for complex decisions, such as autonomous driving, the list of ifs and thens grows. However, when we don’t have enough knowledge on whether to make a decision, we chose to make “satisficing” decisions i.e. we make decisions that are good enough. 

An example of satisficing:

“It’s not intuitive for most people to think of airport lounges as a solution to poor prediction and that they will be less valuable in an era of powerful prediction machines.”

Chapter 11. Fully Automated Decision Making

While the chapters before convinced us that humans have a role, this one gives examples how that won’t be true for long for many decisions.

Two examples: a) how Tesla’s autopilot has averted accidents,  

and b) autonomous trucks to handle the heat in Australia for mining iron.

Chapter 12. Deconstructing Work Flows

Since AI tools can be built to accomplish small tasks, to complete bigger projects engineers will need to identify all the tiny tasks to apply AI tools fully. But if the whole project can’t be completed using AI, users will still see immediate benefits by applying AI to small tasks. This may not be practical, however. The authors provide iPhone’s keyboard as an example of AI completing a small, but an important task from the bigger project.

Chapter 13. Decomposing Decisions

This chapter builds upon the previous chapter by providing a framework to list all different AI tasks required to complete a project or an idea. They call it the AI canvas.

Chapter 14. Job Redesign

This chapter addresses the many concerns people have that AI will replace human workers. It provides an important example of spreadsheets. Spreadsheets dramatically reduced the time it took for bookkeepers to do various calculations, yet they were out of a job. That’s the premise of this chapter: AI will give a hand to many workers and workers will spend time asking better, important questions.

Chapter 15. AI in the C-Suite

I didn’t understand this chapter well. I believe it is saying that the managers needs to realize that data is an asset and people who provide judgement using predictions to help make better decisions are valuable. Still, someone needs to take action on those predictions otherwise there’s no value in those predictions.

As I like to say:

“100% of your analysis that the reader doesn’t take action on is wasted.”

Or

“Interesting is not actionable.”

Chapter 16. When AI Transforms Your Business

AI tools can transform a business, at times, blurring or extending the boundaries of its business objectives. That is a decision that managers need to make: whether to offer services (or data or predictions) using the proprietary data that they have or to collect more data for higher efficiency.  

“Since judgement likely to be the key role for human labor as AI diffuses, in-house employment will rise and contracting out labor will fall.”

Chapter 17. Your Learning Strategy

This chapter outlines various ways AI can be used and implemented i.e. supervised vs unsupervised learning, simulation, and cloud vs local. It also provides useful information to business leaders who are thinking of using AI tools for their businesses.

“The benefit of deploying [AI tools] earlier is faster learning, and the cost is greater risk [to the brand or company].”

Chapter 18. Managing AI Risk

This chapter identifies the six main risks that machine learning faces: liability risk, quality risk, security risk, input data risk, training data risk, and feedback data risk. The data ones are obvious: if the machine learns from faulty data, the results would be faulty too. To manage the first three risks, however, leaders need careful planning and implementation. The authors provide Microsoft’s chatbot fail as an example of the data risks.

They offer this advice:

“Balance the trade-off between system-wide risks and the benefit of doing everything a little bit better.”

Chapter 19. Beyond Business

The book ends with this chapter offering thoughts on AI’s impact on society. The authors tackle questions such as, is AI a good thing? Will all jobs end? Will AI create more income inequality?

The biggest difference between prediction machines (what we refer to AI currently) and general artificial intelligence (GAI) is that the typical AI applications learn from given examples (i.e. data) for a specific purpose, but GAI can generalize the learn and do things “generally.” This is known as singularity. We are not there. Yet.

About the Author

The author of Tableau Data Visualization Cookbook and an award winning keynote speaker, Ashutosh R. Nandeshwar is one of the few analytics professionals in the higher education industry who has developed analytical solutions for all stages of the student life cycle (from recruitment to giving). He enjoys speaking about the power of data, as well as ranting about data professionals who chase after “interesting” things. He earned his PhD/MS from West Virginia University and his BEng from Nagpur University, all in industrial engineering. Currently, he is leading the data science, reporting, and prospect development efforts at the University of Southern California.

>