Machine Learning may bring about certain associations to the futuristic vision known from Terminator, however, we are surrounded by this technology today, and it is present in many areas of our everyday lives. It is likely that all of us type in questions into search engines, choose the best routes using GPS or detect spelling mistakes thanks to AutoCorrect. We owe all these facilitations to machine learning. If you have used a recommendation system today, this might mean that the machine has succeeded, learned your preferences and put forward a correct suggestion of an action. How is that possible?
The history of self-learning systems dates back to 1952 and the company called IBM, where Arthur Samuel dedicated nearly a decade to develop a program for training chess players. And it is precisely this company that has made Machine Learning (ML) famous with the loudest chess match of man vs. computer. In 1997, the world champion of the day, Garri Kasparow, was defeated by their Deep Blue computer. Despite many controversies, it was the first time that AI had shown its superiority in such a loud duel. In March 2016, Google skipped further ahead of this success. A computer known as AlpaGo won a tournament with the best player in the world in one of the oldest and most complicated board games to have ever been invented. Go offers a quite larger number of possible moves than in chess, which is why a win in this game has been called a milestone in the growth of artificial intelligence.
AI is a branch of informatics that deals with creating intelligent models of behavior and simulating chosen functions of the human mind. On the other hand, machine learning is an interdisciplinary science that incorporates robotics and statistics that is part of this branch, and deals with algorithms that can self-improve the system’s operation. In other words, they can learn without participation of the coder or improvement of components.
Unlike people, machines do not search for meaning, but they rather look for patterns. The more data they have, the easier they can predict the result. Coming back to your today’s recommendation, self-learning systems constantly monitor our behavior. If the system has previously suggested a film or product that did not meet your taste (lack of target action, such as purchase or click), this is also an important piece of information. Of course, algorithms are not able to check the reasons behind this; they can only verify the probability of the result based on collected information, in other words experiences. Similarly, a smartphone displays information about the density of traffic on your way to work. The system remembers, that Monday through Friday you are likely to be moving towards the address of your firm at a certain time. Many companies are seeing a chance for success in personalization of content, and some have already achieved great success.
The internet of things, which is a concept of a network of devices which mutually exchange information, has led to the formation of big, diverse, and variable sets of data (so-called Big data). Depending on the nature of this information, these may even be petabytes of content. Meanwhile, the processing power of computers is evolving as well; it is precisely these two factors that have contributed to the growing importance of Machine Learning. The Internet offers millions of examples which can be used by computers to learn thanks to parallel processing. Great challenges facing coders are acceleration of learning, decrease of the necessary number of examples and regularization, which is a situation where programs don’t pay attention to irrelevant changes in examples.
Is it worth investing in Machine Learning?
According to a report by Gartner, an independent analytical research company dealing with the issue of strategic use of technology, as little as 15% of corporations have been able to successfully implement machine learning. Many companies are still afraid of investing in these tools. Are they right?
Machine Learning can definitely find a practical use in a smaller business. Of course, this should be preceded by proper classification of the problem and a relevant selection of representative data, so that the system can generate correct behavioral patterns. If we wanted to carry out an analysis of data to pick up on any irregularities, say a client’s profile and the large size of the database wouldn’t allow us to do this with traditional methods, for example in a company where data from many branches is compiled, we could benefit from unsupervised learning. In turn, supervised learning would be of use in a financial institution to accelerate decision-making with simultaneous decrease in the level of risk. In the future, production lines may benefit from reinforcement learning, in the case of which the system performs both correct and incorrect decisions. The first are enhanced in such a way that the machine may operate automatically. Of course, implementation of automated mechanisms of machine learning is no easy task. However, selection of the right algorithms and data for training and constant monitoring of variables may bring fantastic results.
Mobile Deep Learning
Artificial neural networks are computer models designed to mimic the aspects of the structure and function of the human brain, conveying information between the synapses and nervous systems. An example of their use is perceptive tasks, such as recognizing speech of images. Some of the known examples are Google’s voice search, Facebook’s facial recognition in pictures or automatic sorting per content in Google Images. These systems work on powerful servers in the cloud; however, thanks to the introduction of AI accelerators to the market, it has become possible to convey these solutions outside the network. In smartphones, a supporting module like this one relieves the main processor, optimizes the app’s operation and decreases consumption of energy. Currently, most premium phones have neural coprocessors. And this is just the beginning; in the future, AI in mobile devices is supposed to anticipate our needs.
Today, our phone gives us a heads up about the traffic density in the neighborhood, and in some time, manufacturers are planning to expand this function; our smartphone, based on information in the calendar regarding our foreign business trip, would download all the necessary apps and data before we would actually get on the plane. After we would arrive, we’d have all the information on travel to the chosen target at hand, topped with info on public transport and cheap taxis. Communication troubles would also become a thing of the past; the device would use voice recognition and download a language package for machine translation in advance (this is already possible in Google Translate). A device that adapts to the conditions on its own and plays the part of an intelligent assistant is no distant fantasy.
Learning AI in video games
Last year during the electronic entertainment expo in Los Angeles, Electronic Arts, one of the greatest giants of the industry, announced cells specializing in testing and using machine and deep learning in video games. Search for Extraordinary Experiences Division (short for SEED) has recently trained an AI agent in the game Battlefield 1. AI spent the first 30 minutes observing the behavior of average players; using several computers, it spent the next six learning how to play on its own on a specially tailored map. Interestingly enough, in this case the goal is not to create an agent who will have no trouble winning with professional players, but to increase the player’s immersion through the illusion of contact with a live human. Turn 10 was guided by the same idea with their series of car simulations Forza Motorsport. The program analyzes players’ behavior and, on this basis, creates their virtual avatars which can later be battled by other players. At this point, developers stress the difficulties they may come across using neural networks and unpredictable artificial intelligence when designing games. It seems that a natural environment to use machine learning is competitive gameplay; however, in the future, solutions like this may lead to creation of completely new kinds of video games.
Surely, the majority of people have heard of autonomous cars. The greatest corporations in the world have joined the race; these include Tesla, Google, Uber or most automotive brands. For now, it’s still a long way for drivers to be able to completely let go of the steering wheel. Manufacturers of serially produced cars offer three levels of autonomy, or in other words, conditional automation. Sure, with the right conditions and route, the care can run on autopilot, but the driver must react to the system’s notifications. It is not just technical, but also legal aspects that stand in the way of full automation. In Poland, due to lack of proper regulations this option is blocked in cars coming straight from the showroom, which does not mean that the world is not seeing its first attempts to commercialize a car autopilot. In some cities, fully autonomic taxis have already made their way to the streets. Of course, this is not on a mass scale. For example, Waymo, a company from the Alphabet Inc. conglomerate called by Google is teaching its cars behavior on the road in one of the districts of Phoenix, Arizona. Waymo taxis have been equipped with many precise sensors which create a very detailed 3D map of the area. The system analyzes the changing condition of the vehicle, events on the road and chooses the best solutions. Many companies are planning to own such unmanned fleet until 2021.
The tip of the iceberg
In America, Google has recently shared a system of automatic responses in its mailbox client. On the basis of analyses of collected correspondence, Gmail is able to suggest a one-sentence response, e.g., to our supervisor’s question regarding an overdue report. The user may confirm compliance of the response and return to his work. Drones have an advanced video equipment and software which enables tracking of an object with simultaneous avoidance of any obstacles. Netflix has become known for tremendously tailored recommendations of movies and series. This has almost become the foundation of the company’s success. The system collects information not just about the user’s interest in a certain image, but also his pausing or cancellation and return to the menu. The first chatbot called Cleverbot, which has been recognized as humane by 59.3% of respondents has also passed the Turing test.
Development of robots, voice-controlled user interfaces, automatization of production and coal mining systems, recognition of diseases based on symptoms, steering spaceships (NASA Remote Agent) and systematics of astronomical objects (NASA Sky Survey), recognizing writing based on examples, anticipating trends on financial markets and many more. Examples of the already existent machine learning solutions may be multiplied, and this will only become the tip of the iceberg, which is just revealing its true face.
The problems of machine learning
The algorithm of deep learning in a very simple way enables manipulation of an imagine in video recording using free tools. This includes the open source programming library made available by Google in good faith - TensorFlow, or the Face2Face algorithm which can modify facial expressions in real time. This may lead to misuse, such as fabrication of obscene pornographic films with famous persons.
Machine learning systems are not perfect yet. They are used for so-called attacks on neural networks. The Internet is loaded with images contaminated with noise invisible to humans, which makes programs identify an object erroneously. And so, the image of a dog may be interpreted, for example, as a dolphin etc. In the future, AI may be used for automatized hacker attacks, which creates an entirely new threat.
The machine learning of tomorrow
Computers still have a lot to learn. Machine learning is the fastest developing branch of artificial intelligence. It is estimated that till 2025, investments in this area of informatics will exceed the value of 100 billion dollars. Google is at the forefront of research; today, it offers a range of services using its possibilities: recognizing speech and images (including written language), neural machine translation, thanks to which the translator interprets whole sentences according to context, and not word by word etc. But it’s not just Google that is seeing the unused potential of this solution. More and more companies from completely different fields and scale of operation are dedicating their means for development of machine learning tools, such as Netflix or the Polish Allegro. From elastic anti-spam filters, trough automatic responses and personalized service, ending with the industry or finances – in the future, machine learning may change the way people think.