Machine Learning is a branch of artificial intelligence that allows machines to learn without being expressly programmed. An essential skill for making systems capable of identifying patterns among data to make predictions. This technology is present in many applications such as recommendations from Netflix or Spotify, intelligent responses from Gmail, or speech from Siri and Alexa.
“In short, ‘machine learning’ is a master of pattern recognition. And it is capable of converting a data sample into a computer program capable of drawing inferences from new data sets for which it has not stood previously trained,” he gives details. This learning capacity remains also used for search engine improvement, robotics, medical diagnosis. Or even the detection of fraud in the use of credit cards.
Although fashionable now, thanks to its ability to defeat Go players or solve Rubik’s cubes, its origin dates back to the last century. “Statistics is undoubtedly the fundamental basis of machine learning, which consists of a series of algorithms. Capable of analyzing large amounts of data to deduce which is the most optimal result for a given problem,” adds Espinoza.
Table of ContentsToggle
Old Math, New Computing
You have to go back to the 19th century to find some of the mathematical milestones that laid the foundations for this technology. The Bayes ‘ theorem (1812) defined the probability of an event occurring based on knowledge of the preconditions associated with this event.
Years later (in the 1940s), another series of scientists laid the foundations for computer programming. Capable of translating a series of instructions into actions executed by a computer. These precedents made it probable for the mathematician Alan Turing to pose for the first time in 1950 whether machines can think. He planted the seed of the creation of ‘artificial intelligence’ computers. Computers capable of autonomously replicating typically human tasks, such as writing or image recognition.
A little later, between the 1950s and 1960s, different scientists began to investigate how to apply the human brain’s neural networks’ biology to create the first intelligent machines. The idea led to creating artificial neural networks. A computational model inspired by how neurons transmit information through a web of interconnected nodes. One of the first research questions was Marvin Minksy and Dean Edmonds, scientists at the Massachusetts Institute of Technology (MIT). Both managed to create a computer program capable of learning from experience to get out of a maze.
Machine Learning is a Master of Pattern Recognition
it programmed to do so but only did it by learning from the examples provided initially. The achievement signified a paradigm shift from the broader concept of artificial intelligence. “The great milestone of ‘machine learning’ is that it allowed us to go from programming using rules to letting the model bring out these rules in an unassisted way thanks to the data,” explains Juan Murillo, Manager of Data Strategy at BBVA.
Despite the success of the experiment, the achievement also highlighted the limits that technology had at that time. The lack of data availability and the lack of computing power meant that these systems did not have enough capacity to solve complex problems. It led to the influx of the so-called ‘ first winter of artificial intelligence, a series of decades during which the lack of results and advances made the academic world lose hope regarding this discipline.
The Rebirth of AI
The landscape began to change at the end of the 20th century. With the advent of the internet, the massive amounts of information available to train models increase computers’ computing power. “Now we can do the same as before, but a billion times faster. The algorithms can test 500,000 million times the same combination of data to give us the optimal result. And they do it in a matter of hours or minutes, whereas before it would proceed weeks or months,” says Espinoza.
In 1997, a well-known milestone marked the rebirth of machine learning: the IBM Deep Blue system, trained by watching thousands of successful chess games, managed to defeat the top world champion of this game, Garry Kasparov. The achievement was possible thanks to ‘deep learning’ or deep learning, a subcategory of ‘machine learning’ first described in 1960, which allows systems to learn from experience and train themselves to do better and better using data.
The difference with previous techniques is in its ability to adapt to changes in the data as they enter the system
This technology has advanced exponentially over the last 20 years and is also responsible for AlphaGo. The program capable of defeating Go’s game to any human player. And what is more critical: to train by continually playing against himself to keep improving.
The system used by AlphaGo to achieve this is, precisely, ‘reinforcement learning.’ Which is one of the three main currents that remain currently used to train these models:
- ‘Reinforcement learning’ occurs when a machine learns through trial and error until it reaches the best way to complete a given task. For example, Microsoft uses this method in-game surroundings like Minecraft to see how ‘software agents’ improve their work. The system learns to modify its behavior based on “rewards” to solve the assigned task without explicitly programming it to perform it in a certain way.
- ‘Supervised learning’ occurs when machines remain trained with tagged data. For example, photos with descriptions of the items that appear in them. The algorithm used by the device is capable of selecting these labels from other databases. Thus, if a group of images showing dogs has stood tagged, the device can identify similar images.
- Finally, in the case of ‘unsupervised learning,’ machines do not identify patterns in labeled databases but instead look for similarities. In this case, the algorithms remain not programmed to detect a specific type of data. As was the case with dogs’ images, but instead, look for examples that look alike and can remain grouped. For instance, in the case of facial recognition, it is what happens. In which the algorithm does not look for specific features, but rather a series of common patterns that ‘tell’ it that it is the same face.
Flexibility, Adaptation, and Creativity
The models of ‘machine learning’ and specifically reinforcement learning or ‘reinforcement learning’. Have a characteristic that makes them especially useful for the business world. “It is its flexibility and its ability to adapt to changes in the data. As they enter the system and learn from the model’s actions. Therein lie the learning and dynamism that the previous techniques lacked,” adds Juan Murillo.
Now the case of AlphaGo, this means that the mechanism familiarizes according to the opponent’s movements. And uses these new figures to continuously improve the model. The latest form of this computer, called AlphaGo Zero. Can accumulate thousands of years of human knowledge in just a few days of operation. Also, “AlphaGo Zero also discovered new knowledge, developing new creative strategies. And unconventional movements,” explains DeepMind, the Google subsidiary company responsible for its development, in an article.
This unmatched compliance and inventiveness have enormous potential in the future to improve scientific disciplines. As diverse as the creation of synthetic proteins or the design of more efficient antennas. “Industrial applications of this technique include constantly optimizing any ‘system,'” explains José Antonio Rodríguez, a senior data scientist at BBVA’s AI Factory. Also, in the world of banking, deep learning allows. “Creating algorithms that can adjust to changes in the behavior of markets. And customers, to balance supply and demand, for example, offering personalized prices,” concludes Rodríguez.
Improvement of Autonomous Cars
Another example is the improvement of autonomous cars, which have made great strides in recent years thanks to ‘deep learning. It allows them to improve their precision progressively. The more they drive, and the more data they can analyze. The possibilities of ‘machine learning’ are virtually endless as long as there is data available to learn. Some researchers test the limits of what we call creativity, using this technology to make art or write articles.