How Do Machines Learn? Part I: They Train

Machine learning part 1

Photo by Robert Couse-Baker. Click image to see his photos on Flickr. Attribution 2.0 Generic (CC BY 2.0). The image has been modified with text.

Machine Learning and Artificial Intelligence are common buzzwords that seem to be everywhere in 2019. Every corner of society is affected by the rise of the machines along with a sense of wonder and fear as to what this means for humanity. Will the machines take all of our jobs? Can machines really think, and at what point do we have machines that are smarter than humans? If science fiction movies are to be believed, then the only eventual outcome is war. In separating fact from fiction, this is the first in a three-part series for On the Brink where we explore some of the big ideas in machine learning.

Is Skynet Already Here?

It would seem like the first immediate question is whether artificial intelligence already exists? If the term is banded about so heavily, does it mean the machines are already thinking (and outthinking) human beings? Not exactly. In the realm of computer technology, data science, and mathematics, when we use the term Artificial Intelligence we are really referring to a set of processes that computers use that mimic or replicate cognitive thinking commonly associated with humans and other animals that may be involved with complex problem solving or some complicated or intricate task. The computer algorithms designed to develop machines that cut automotive parts to detailed specifications are (on the whole) much less complex than what would be needed for a machine to drive a car or to compete as a boxer or to learn to dance. Science has not reached the point where machines perform sentient and independent thought, and the terms machine learning and artificial intelligence still only apply to those computer algorithms that can do these higher-order tasks. Indeed, we are still ourselves learning about what machines can and can’t learn to do.

Machines Need Lots of Data

So is all of this simply about developing increasingly complex computer code? Can a programmer just sit down and write millions of lines of code so that a computer can respond to every scenario? No. Recent innovations have allowed us to create algorithms that can use data to gain understanding on how specific tasks are performed. Machine learning took off as part of the era of Big Data and it is this increased access and availability of data that has helped spark this revolution. We use these large amounts of data to train machines to do a variety of tasks, things that mostly involve making complex associations between many different factors simultaneously. These algorithms are as good, if not better, than humans at making these kinds of multidimensional connections. They can also “play out” many different possible scenarios more quickly than humans. This means that these algorithms can now perform some complex tasks that were previously relegated to human specialists. Here we aren’t just talking about playing chess or Jeopardy but also in searching medical images for cancer detection, identifying fraud in financial systems, preparing air-tight legal documentation, or determining the authorship of famous historical documents.  

Data feeds these computers to fine tune increasingly complex mathematical algorithms that the machines use to predict known events within the data they have been given. Once the machines adequately predict on the training data, those algorithms are deployed out into real world settings to help perform those tasks in settings where new data is currently being gathered. So the secret is that a machine has to be shown medical images where cancer has been identified so that it can develop a mathematical model to apply to medical images where no one has yet identified whether there is cancer or not. If a machine can scan those images as well as a trained physician, then it can scan images faster and without the need for rest or sleep.

So how do machines get the data on medical images where cancer has been identified? Currently, it has to be generated by experts (ie, humans). Machine learning still needs human teachers.

Boxing, Dancing, and Driving

So how would we train a machine to do very complex human tasks such as boxing, driving, or dancing? These kinds of tasks have proven increasingly difficult and need lots of data from many different aspects of the task to be studied. Driving not only requires an understanding of road conditions and how to follow the lines on the road across complex terrain but also weather conditions, road hazards, traffic patterns, lights, and signs. An excellent boxer performs a slew of different exercises that not only include cardiovascular exercises for agility and endurance but also form work, sparring, strength training, and punching work on both heavy and speed bags. The point is that the human brain takes all this data in and internalizes it when we learn how to do these things. Humans have the benefit of learning even when we aren’t training to perform certain tasks. A typical American has been a passenger in a car for 15-16 years before ever getting behind the wheel to learn how to drive and will already have insights and experiences on weather, acceleration, and braking long before studying the mechanics of driving.

Machine learning currently has an advantage in that it can input a large volume of information quickly and perform many more calculations per second than can a human. However, humans train with data our brains gather every day. We have an advantage in that we create mental algorithms for dealing with all of our complex tasks, and we routinely repurpose bits from certain algorithms for use in other dealings. To teach a computer to dance, we would have to start from scratch, whereas humans learn to walk and jump well before they learn about rhythm or technique.

Good Teachers

Machine learning algorithms are only as good as the data they are trained with; this means that if the process under study is biased, then these algorithms can amplify those biases. If credit score models are inherently built upon a structure that is discriminatory towards minority groups, then machine learning algorithms can be racist or sexist. Just like humans, machines are not born prejudiced but can act a certain way if that is what they are taught. This is why the need for quality data collected in a well-thought and meaningful way is essential. Here a good teacher can make all the difference, just like with humans.

Tune in next month for part two of this series: How Do Machines Learn? They Fight.

Jason Brinkley bio

Jason S. Brinkley, PhD, MA, MS (Photo: American Institutes for Research)

Jason S. Brinkley, PhD, MS, MA is a Senior Researcher and Biostatistician at Abt Associates Inc. where he works on a wide variety of data for health services, policy, and disparities research. He maintains a research affiliation with the North Carolina Agromedicine Institute and serves on the executive committee for the NC Chapter of the American Statistical Association and the Southeast SAS Users Group. Follow him on Twitter. [Full Bio]

Previous posts in this series: