Learning seems to be an important aspect of many AI technologies, but not all AI can learn. There are some computers considered as AI which have been pre-programmed with all the behaviors they have, with no ability to learn or do anything else, and then there are some that can learn. Today let’s focus on the ones that learn, and on some specific ways how they do it. Machine learning is the general term for how machines are able to learn and improve from experience without being specifically taught and programmed for each task. It’s based on big collections of data that the machine can access and use to learn from them. Two very interesting sub-sections of machine learning that are worth a mention are deep learning and reinforcement learning.
Deep learning is one of the most widely used machine learning methods. It uses algorithms inspired by the human brain – called artificial neural networks. It’s the base of many AI technologies we have or are in the process of developing, such as driverless cars, face recognition, speech translation, virtual assistants and voice control on different devices. It is also used for human safety in work environments with heavy machinery. Some of the most important uses have been in medicine, helping recognize cardiac arrest cases in emergency calls or improving medical diagnoses based on the analysis of medical images like x-rays. The use of deep neural networks works well in medical diagnosis because the input criteria that it uses for illnesses, tumors and fractures are often well defined and pretty narrow, which is important in order to ensure a high quality of results. Deep learning hopes to replicate the way human brains naturally learn and function, through the use of large amounts of labeled data and neural networks composed of many different “layers” which communicate with each other. These artificial neural networks mimic the fundamental functioning of the human brain: our neurons and neuron connections, and the way our brains learn by modifying these connections. It requires quite a lot of computing power, which is why even though it was first theorized in the 1980’s, it wasn’t a realistic possibility until a while after, when this level of computing was achieved. Deep learning is important because it has led to a much higher level of accuracy in many of its uses, where the AI can even outperform humans in some tasks like image recognition and classifying, and this will probably allow for much more sophisticated advancements and uses as it evolves.
Reinforcement learning is a subfield of complex machine learning algorithms mostly used in robotics. It is modeled on the study of animal behaviour, where learning is based on a lot of trial and error. The intelligent agent is driven by a goal, and is rewarded or punished based on the quality of the actions it takes. The “good” actions – ones that lead to successful results – have a better probability of being repeated. In reinforcement learning it is important to have a list of available actions, and a way to measure if the outcome was successful or not, after which the agent uses trial and error to learn and improve over time. Besides in robotics, it is also used for games where the agent can try different scenarios and based on the result and feedback adjust the play strategy and the moves it would choose in the future. The interesting thing is that the rules of the game are programmed into the machine, but without any input on good or bad options and solutions. Through completely random attempts in the beginning, the machine reaches very accurate and sophisticated skills in the end, based on a lot of attempts and the goal to maximize the total reward. The tricky thing with reinforcement learning is providing the simulated environment for trial and error learning. With Chess or similar games, it’s fairly simple and easy to program the rules and environment in which the intelligent agent learns, but if we take a reinforcement learning approach to building an autonomous vehicle, it can get quite complicated. And even if a training simulator was built and the vehicle could learn how to drive well according to laws and rules and avoid collisions, it’s not so easily transferable to the real world.
Deep learning and reinforcement learning are just two important subfields within machine learning. It’s important not to regard either of them completely separately, from machine learning but also from each other. Although the idea and methodology differ, they can be used in combination with each other in different projects, in order to accomplish more complicated technologies, such as autonomous vehicles. It’s important to be aware of their differences in the way the intelligent agent is trained and how it leans, but also to allow for the possibility of interconnected technologies which could be improved by using more than one method of machine learning in order to accomplish more sophisticated results.