Demystifying machine learning: How do machines really learn?
It seems that everybody is talking about artificial intelligence and machine learning these days. But machine learning is not a new concept — it’s actually been around since the late 1950s. So why all the hype? Why do businesses feel like now is the time to adopt? Before businesses start to develop a strategy around machine learning and AI, it’s important to review how machines really learn, and how this can impact your AI and machine learning strategies.
To start, there’s been a lot of discussion around the difference between AI and machine learning, and even the term analytics has become very nebulous. What’s important from a business perspective is not the technical definitions, but the value it brings you. All of these technologies are powered by data, and the goal for a business is simply to do what it takes to move from data, to insight to outcomes as efficiently as possible.
One area where there is a clear connection between data, insight and outcomes is in the field of maintenance. People have long relied on routine maintenance as an attempt to ward off potential problems. Until recently, we changed our oil every 5,000 miles. Not because there was an indication of failure, but because we were too reliant on routine. In short, we didn’t have the data to tell us whether we needed to change the oil. In this case, it wasn’t a major problem since it’s fairly inexpensive to get your oil changed, but imagine if you had thousands of machines or extremely expensive machines? Taking a single expensive machine offline, like an aircraft engine, when it’s not necessary or managing thousands of machines using a static maintenance routine are both problematic. Read: You’re wasting money when you’re making maintenance changes based on routine. Instead, businesses need to start using data to move away from routine maintenance, which will help save money in the long run.
The large volume of data presents a tremendous opportunity, but it can be hard to take full advantage. After all, you don’t simply need more data to guide your decisions — you’re likely drowning in that already. In fact, one of the biggest questions that you may be facing as a business leader is how to utilize the data you’re already generating on a daily basis.
Perhaps you’re already employing data scientists to build the models you need to make sense of it all, but they can only process a fraction of the data that is flooding in. Hiring more data scientists is not only expensive, but there is such a shortage that hiring enough may be a logistical impossibility. A better solution is to utilize a machine learning platform to remove this burden from the team processing your data, allowing them to focus on making decisions while the platform focuses on analysis at scale.
As we covered above, more data is readily available than ever before. Beyond the IoT sensors that constantly generate machine data, there’s also behavior data from customers and users, and business transaction data in an ERP or CRM itching to be analyzed. Data is in such abundance that the level of contextual insight that can be derived is phenomenal.
Humans, however, have a hard time processing data at this level. The average human is capable of processing several dimensions at once — but when too many variables are introduced, even the most intelligent among us can’t process it all. We start eliminating variables, or we generalize, which reduces the accuracy of our analysis.
Machines, on the other hand, do not have this limitation. Computers boil everything down to binary decisions — yes or no states, 0s and 1s — and eliminate nothing. Mathematical models can then be designed to simulate real-world behaviors, and the computer can run the entirety of the data input against the model to yield a result that is as accurate as possible.
The key then to producing accurate results lies in the model. Ordinarily the creation — and critically, the frequent tuning — of a model is a complex task, requiring the painstaking labor of a team of trained data scientists. To keep the model up to date, it must take in the very latest and most relevant data, and the new analysis must be compared with past analysis and results, leading the data scientist to a series of conclusions that allow them to further increase the accuracy. This can be a difficult and time-consuming process.
Fortunately, it’s all math, and we can use machine learning to automate much of this process — I know it sounds strange, but we are basically using machine learning to automate the machine learning process! Why rely on data scientists to do the manual work of running and deciphering which set of algorithms provides the best prediction? A computer can automate the testing of the models without limiting the tests to a subset of the data. And a computer can do this continuously in order to respond to operational or environmental changes. While it appears as if the machine is learning, it is merely improving the accuracy of the models that will generate the most accurate result. This approach allows you to use your data scientists to analyze the outcomes versus spending all their time doing routine work. Or to enable the average business user (like me) to make sense of the predictions, even when we don’t understand the math. So, while there’s no actual “learning,” the magic is in the math, which is now available to the average organization.
All IoT Agenda network contributors are responsible for the content and accuracy of their posts. Opinions are of the writers and do not necessarily convey the thoughts of IoT Agenda.