Two words that are spoken in every leadership and board meeting around the world right now are “machine learning”. Technology buzzwords seem to monopolize these meetings. Who could forget: digital, big data, internet of things (IoT), mobility, …-as-a-service, security, the cloud and the recent favorite, blockchain? Now, machine learning, deep learning, reinforcement learning, and numerous other technological terms that describe the artificial intelligence space have become this year’s buzzwords.
I’ve been in meetings with other executives where most people, including me, can’t make heads or tails of what people are talking about when this subject comes up. It’s like listening to a foreign language. If you are a CEO, your closest confidants are throwing all these words at you and you are thinking to yourself: “How did they learn this concept so fast? When did they become the expert in this space?” Most of them are just serving up the alphabet soup of the latest buzzwords to mask their own ignorance. It’s up to you to figure out who knows what they are talking about (“contenders”), who doesn’t (“pretenders”), and whether this machine learning technology really matters to your business.
I have been trying to sort this out for the last three years; almost full-time since I retired from my post as CEO of Accenture Operations and started my role as Managing Director at Carrick Capital Partners, where I am focused on investing in the ML space. The reason this is hard to understand is that ML is a combination of mathematics, computer science, and statistics. These are highly technical areas and usually come with their own vocabulary. Most ML scientists want to do research and make groundbreaking discoveries they can publish. They do not want to spend their time explaining these definitions and teaching business people how to apply existing ML technology to their businesses.
In working closely with ML scientists, I have found that there are countless discarded pieces of ML technologies sitting around university labs because they are no longer needed or relevant to specific scientific research. But, to CEOs who can understand and identify their potential, these “gold nuggets” have enormous potential to significantly increase their revenues or decrease their costs.
Now, decades after scientists started working in the space, businesses are recognizing the potential and want to align with experts to bring some of these “gold nuggets” from the research lab into the business environment. The first step in this journey is for CEO’s to take the time to understand the terminology so they can avoid the “pretenders” and quickly engage with the “contenders” who can help them deliver real business impact in the area of ML. What follows is a “kitchen English” cheat sheet of definitions for CEOs with anecdotes about how they are relevant to business. Please save them in the notes app of your smartphone for your next machine learning discussion.
There are a variety of acceptable definitions for AI, but generally speaking, it is a term used to describe when computer science solves cognitive problems. More simply put, computers doing work that would otherwise require human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages. As a CEO you probably already know that the term AI makes people nervous. People performing transaction work will fear for their jobs and your HR department will claim that they need to radically change their approach to training. Don’t let the hype worry you or slow your progress. I know from firsthand experience that the job of getting your data prepared and clean so it can be used for AI will provide more than enough job opportunities for your workforce.
“ before you spend a lot of money, make sure the algorithms that you are using are designed to accomplish a task that can have business impact.”
ML seeks to use data (input) to uncover relationships and make predictions (output). The computer (“machine”) “learns” relationships in data, including learning to make predictions. Replicating human cognition is not an explicit goal of ML (as it often is in AI, at least historically), and ML can make predictions that would be hard for any one person. As an example, an ML algorithm can learn to predict health or disease by analyzing all of the data generated by medical specialists, each of whom may see the patient from only a single perspective. The “machine” can see patients from a much more complete perspective by analyzing all of the available data.
An algorithm is a set of guidelines that describe how to perform a task. It can be as simple as a recipe or driving directions, or as complex as a sequence of instructions telling a computer what to do. With AlphaGo, an algorithm was created to play the complex Chinese game called “Go.” The experiment was a success, but there was no obvious business application or impact. The takeaway is that before you spend a lot of money, make sure the algorithms that you are using are designed to accomplish a task that can have business impact unless you just want to do a science experiment.
Neural networks were originally developed in the 1980s, with the goal of emulating the processing capabilities of the human brain. While there are connections between neural network algorithms and the way the brain works, there has been far less recent emphasis on connections to the brain, and more emphasis on developing neural network algorithms that work well in practice. Neural networks have had a long history, characterized by over-hype, failure, rebirth, hype, failure, and on and on. This process has taken place for more than three decades.
After 2010, two important things came together: access to massive digital data and significantly enhanced computing power (e.g., graphical processing units, or GPUs). Together, they have pushed neural networks back into a period of frenzied hype, which is why you are hearing about them in every board room and leadership meeting. The neural networks that work well have many layers, which make them “deep.” The deep neural networks used to perform ML have led to the term “deep learning.”
When supervised learning is performed, both inputs and outputs are known and labeled. Imagine you are translating English to French. If the input is “Yes” the output is “Oui.” Right now 99% of the economic value that is derived today is due to supervised learning. This is the most accessible form of ML which is most commonly used in business. However, supervised learning can be expensive to do at scale because labeling data is often a manually intensive task.
This is the ability for technology to learn to figure things out independently from unlabeled data. In one example, ML was programmed to watch YouTube and figured out the concept of cats by itself. This is the Holy Grail for business because it requires less labelled data, which significantly reduces the need for human labor.
Semi-supervised learning falls between unsupervised learning and supervised learning in that it is able to make use of unlabeled data by also leveraging a small amount of labeled data. This allows researchers to produce improved accuracy. Due to the expense associated with the human labor required to label data, semi-supervised learning is a cost effective approach. It leverages the inexpensive un-labeled data to achieve greater accuracy than unsupervised learning, and uses less labeled data than supervised learning. For businesses only able to perform supervised learning, this offers a “training wheels” approach to move toward unsupervised learning.
With transfer learning, you use data to solve one problem, such as how to recognize cats and dogs, and use what you learned to solve a different problem, such as recognizing other types of animals. This is most commonly applied where you have lots of labeled data in one area (e.g., cats and dogs), but little labeled data in another area (e.g., horses).
The human learning process is experiential. You first perform an action, assess the response to the action (good or bad), then tailor subsequent actions to optimize for good results and avoid bad results. In reinforcement learning, an algorithm learns to build a “policy” in the same way. The algorithm sequentially gets data as input, over time. The algorithm dictates an action based on input data, and then the system is given a reward (or absence of reward). Over time, the algorithm learns in a data-driven manner to create a policy such that with new input data, the system selects the best next action.
As mentioned, I have been working on business applications for machine learning for years and I will admit I still don’t have everything figured out, but I am getting closer. My trick was to work with “contenders” like Dr. Lawrence Carin, Duke University’s Vice Provost for Research, and Professor of Electrical and Computer Engineering; his uniquely qualified research team of machine learning experts; and Robbie Allen, the past CEO of Automated Insights, author of multiple technical books and owner of numerous patents.
If you are struggling in the world of ML, the quickest way to accelerate your education is to learn these terms and decode the alphabet soup of this space. Equipped with a better understanding of ML terminology, you will be able to identify the “pretenders” and surround yourself, like I did, with “contenders”, which will allow you to drive business impact through machine learning.