When we talk about machine learning, we mean, in general, the ability of certain machines and, therefore, the computer systems that govern them, to “learn” from experience. Modern computing is based on algorithms (mathematical formulas with many factors) that can become very complex, and on “decision trees” containing instructions for decision-making of the type “if X reaches a certain value do this” or “if A happens, do B” (said in a tremendously simplified way).
Machine learning techniques allow these algorithms to not only take into account the input data it receives, but analyze whether the final result is appropriate (according to previously set parameters) and take those results into account when the algorithm reruns. It is, therefore, a type of artificial intelligence.
Until not so long ago, machines were limited to executing the instructions given to them regardless of the final result. That is, in the face of a mismatch it was a human operator who had to intervene, either in the machine itself or in its programming, to correct the problem. At most, for example, industrial robots were prepared to stop if an error occurred or there was a failure in the data or input components. That is, a soldering robot in an automobile production chain was programmed to weld specific points previously programmed, but it was not able to know if the weld was right or wrong. At best, they could stop (and with them the whole chain) if they detected that the body to be welded was misplaced. But they weren’t able to learn.
Machine learning-based systems are designed to learn in three different ways:
- Supervised learning: In this case there is a “training” supervised by a human operator. For example, if I want a robot to distinguish between dogs and cats, I will have to provide you with the data to help you identify what a dog is and what a cat is, and tell it by means of “tags”. Example: Showing you images of many different dogs and cats, each labeled “dog” or “cat”, so that when the system has enough data it can recognize a dog or cat by itself for similarity.
- Unsupervised learning: In an unsupervised learning model, following the example above, we provide the system with photos of dogs and cats, but we don’t tell you anything about them. The system learns to recognize that there are two different categories (although you won’t know what they’re called) and will be able to, when we show you a photo you’ve never seen, tell us if it’s A or B. If the system is able to connect to different external databases (e.g. the internet), it will look for similarities with the images you have seen, words related to them and end up naming them “dog” and “cat” without anyone telling you.
- Reinforcement learning: The system learns from experience, not only from the data provided to it (input data) but from whether its decisions have been accurate or wrong (output data). Wrong decisions “penalize” the system not to repeat them, while the right decisions reinforce it.
The role of Machine Learning in Industry 4.0
The ability to learn is not only already present in industrial robots, but in many computer applications: Google and Facebook are a good example of how, through continuous exposure to user searches, preferences and interactions, they are able to to make a “robot portrait” of each of us and our tastes. Although we don’t tell you, they know our age, our sex, what we like to eat, drink, our favorite trips, who we like and who we don’t and hundreds (if not thousands) of other data that are used to show us, both in search results and in the advertising, what we’re really interested in seeing.
Obviously, Google or Facebook algorithms are extremely complex and require a mireal computational power. However, industrial processes (both in manufacturing and services) are developed in much more controlled environments, where the input data is not as numerous. This simplifies algorithms and decision trees and requires much less computing power, making them accessible to many companies in their everyday applications. For example, a robotic warehouse is already able to decide, depending on the dimensions and volume of the object to be stored, the most efficient way to do so taking up as little space as possible. This allows more products to be stored in less space, something that is already applied not only by factories, but by retail establishments that handle many different references, such as pharmacies.
Sensors are the key
An automatic Machine Learning-based system cannot learn if you do not have input data. Obviously, it’s no use having an intelligent system if we have to feed it data manually. To do this, industrial automation requires sensors of very different types: cameras that allow to recognize shapes and objects, location and position sensors in space, pressure sensors, temperature, lidar sensors (for example in autonomous vehicles) and a long etcetera. These sensors are the “senses” of a Machine Learning-based system.
All these auxiliary systems that enable the proper operation of an intelligent system are not only that they are already available, but have experienced a very significant reduction in their cost in recent years. This, along with advances in connectivity, makes the application of Machine Learning already possible outside the select circle of large corporations.
What benefits do these systems bring to a company?
Many and very diverse. Compared to traditional automatic systems, Machine Learning-based systems have proven to be more versatile and adaptable, dramatically reduce the error rate, are able to interact in environments where there are humans and each other, and, above all, are able to efficiently perform tasks that were unthinkable for a robot only a few years ago. All this makes them a key part of the productivity increase resulting from the implementation of Industry 4.0.