Module: PMing with AI and machine learning. Lesson: the AI-ML-powered PM. In this lesson, you will learn the differences between artificial intelligence, machine learning, and deep learning, and examples of how these technologies supercharge product managers. AWIP's experience of real PM interviewing indicates that artificial intelligence and machine learning interview questions are generally not asked of PMs. However, AWIP expects interest in this field to increase. We dedicate this time and content to prepare you to demonstrate a firm understanding of high-level AI and ML concepts. Intelligence is the ability to acquire and apply knowledge and skills. For example, think back to when you first learned to ride a bike or drive a car. Compare that to how well you can ride and drive now. What changed? Between when you first started learning and today, you got a lot better, smoother, and more efficient. You can ride and drive with one hand, or for short period of time, no hands. You know how far out to scan the road and when to snap to attention because you saw something unusual. You prefer certain routes, because in your experience, they're better than the alternatives. You might even be able to pull off wheelies or donuts. What I just described is your human ability to acquire and apply natural intelligence. Can machines learn to drive cars in a similar manner? This question spurred research as early as the 1940's into an electronic brain represented by zeros and ones. One famous computer science theory, the Clarke-Turing theory, postulates computers can reason through any problem that humans can and arrive at the right solution. And, in theory, if you had a gigantic enough decision tree with countless if statements, then you should arrive at a car that drives itself. The problem, however, Is that software engineers can only write a finite amount of code, and only account for a finite amount of variables a driver might encounter. The decision tree type of artificial intelligence does not scale, and that's why we didn't have self-driving cars in the 1940s. But by the 1980s, artificial intelligence did get us the ghosts in the video game Pac-Man. These ghosts chased Pac-Man around until he bite a big pellet, and then they turned blue, and ran away from him. Around this time, researchers made progress with machine learning. Rather than program each branch of a giant decision tree, you give the computer a set of inputs, let it generate the predicted output, and then you give it a set of corrected outputs. Then you let the computer adjust its algorithm to reduce the difference between its output and the correct output. Next, you repeat the process with a different set of inputs and correct outputs. Each repetition is called an epoch. Machine-learning algorithms generally need many, many epochs to calibrate themselves. Meaning they need a lot of input. A related field, data mining, is the art and science of acquiring enough, good quality data to supply this volume of input. There are different algorithms for machine learning, but the most accurate and widely accepted is called deep learning. In deep learning, the computer has one or more hidden layers of nodes between your inputs and the computer's outputs. These are the deep nodes. The computer adjusts the weight of each deep node to give the most accurate answers for each epoch. I'll use the terms "hidden layer" and "deep layer" interchangeably. Once you give a deep learning algorithm input, your computer feeds forward your input through the hidden layers of nodes, which make various calculations on the input. Then, the deep layers forward those calculations to the output layer, which for self-driving car, might be, apply brakes or accelerate. Then, when you give the computer the correct output, the algorithm detects any differences and backpropagates those differences to give the correct result next time, or at least to minimize the error and minimize the r-squared. Next, you run the deep learning algorithm with more data as you repeat for many, many epochs, and to develop a finely tuned algorithm for your business purposes. If you've heard of Google's machine learning platform, TensorFlow, the process we described is literally what TensorFlow describes. Tensors are multi-dimensional arrays that represent multiple deep layers of nodes. Flow refers to the feed-forward and backpropagation of information into those tensors. Here are several suggestions for how to apply deep learning in your work as a PM. One, think about customer personas and cohort analysis. In the regression analysis section, I taught you how to find the correlation, but you had to pick a specific independent variable for each experiment. You can use a type of deep learning called unstructured learning to ingest all your customer data. It will try to cluster customers of your customer base by their most relevant characteristics. It tries to pick the most important independent variables that separate, say, a paying customer from a non-paying user. This can help you discover new independent variables to build personas in cohorts around. Two, now, think about A/B/n and tests. Remember Marissa Meyer's, 41 Shades of Blue? During that experiment, she tested the best color for Gmail ads. What if you could also test the word choice, length, position on the Web page, and any number of other variables? You can do this with the type of deep learning called reinforcement learning. Give your algorithm all your variables, let it present those variables randomly to users, and give the algorithm positive reinforcement when a user clicks an ad. Over time and enough interactions, the deep learning algorithm will decide which words, length, position and so on drive the most clicks per ad. In fact, PM software company HubSpot is working on "set and forget AB testing" that does exactly this in the background for you. Discussion: Research how products in your industry of interest, are using artificial intelligence, machine learning, and deep learning. What are the most interesting applications? Do some companies seem to be more aggressive pursuing this field than others? Post a brief write-up of your findings. Two paragraphs maximum on our discussion board.
