An operating system that can communicate with humans may seem like something from the distant future, but thanks to advances in machine learning, it’s becoming more and more of a reality. As machines become smarter by analyzing data and learning through supervised and unsupervised learning methods, they are contributing to solving various problems in our daily lives.
In the movie “Her,” there is an operating system (OS) voiced by Amanda Seyfried. The OS behaves as if it has emotions, storing word-for-word conversations with a person and analyzing them to extract his/her hobbies, tastes, and characteristics. This leads to interesting conversation topics and empathy, and many people, including the protagonist, Theodore, fall in love with the computer operating system. You might think that this computer operating system is something out of a science fiction movie. This is because the machine learns like a human through conversations and lots of experience, and takes new actions based on that. But these learning machines are already all around us. The theory behind these machines is collectively known as machine learning.
What is machine learning?
Machine learning is exactly what it sounds like: machines learning to teach themselves, specifically by analyzing and learning from different types of data that they have acquired in some way, and then providing a mathematical foundation that can be applied to new types of data. The key to this is that you don’t have to explicitly program a computer to learn, you just have to give it input. For example, if you teach a computer to recognize a picture of an apple, it will recognize that it is an apple if you show it a picture of an apple that is different from the previous one, even if you don’t specify the characteristics of an apple.
Machine learning can be divided into two main categories based on how it learns. One is supervised learning, where the outcome (or label) for the data being trained is known. For example, when you post a photo on Facebook, it automatically recognizes your face and displays it as a square. This is done by learning which images are faces and which images are not faces. In other words, if the data is an image file consisting of RGB values of pixels, the result is whether it is a face or not.
On the other hand, another method of machine learning, called unsupervised learning, involves feeding an image into a machine and letting it run without specifying how it should be classified. In the face recognition program above, the machine is able to find features in the image and distinguish whether it’s a human or a cat on its own, despite not being told what to do with the image, i.e., whether it’s a face or something else. Another example is a news categorization program on a portal site. If you look at news articles on a portal site, you may notice that they are categorized into different areas, such as politics, life, entertainment, and so on, and that similar news are grouped together and recommended. To do this, the portal administrator does not specify to the learning machine which articles are political, which are entertainment, which are sports, and so on, but when a large number of articles of various types are fed into the learning machine, it analyzes the frequency of words in the articles by itself and makes a classification with high accuracy for new articles.
Supervised learning methods
The learning method currently used for most machine learning is supervised learning. It accounts for about 95% of all learning machines. One reason for this is that it requires less training data to achieve a certain level of performance than an autonomous learning method, since a human is directly specifying the results for the data. This is because autonomous learning methods need more data to find the characteristics within this data that can distinguish between objects on their own – for example, between a human and a cat, between a political article and an entertainment article. Supervised learning methods, on the other hand, can perform well with relatively small amounts of data because they are guided by humans, reducing the time it takes to collect data and the time it takes to run the training itself.
However, supervised learning has a major drawback: it can’t learn about things that are not taught to it by humans, and its performance can drop dramatically in environments that are slightly different from those taught to it by humans. For example, a machine trained on an image of a person in a bright environment may not recognize the person in uneven lighting or low light, or it may not recognize the side of a person when only the front of the face is input.
Autonomous learning methods
Self-learning methods can address the shortcomings and limitations of supervised learning. However, this learning method reached its limits decades ago. The data required was too large, the computational complexity was too high, the hardware was not capable of handling it, and it was too time-consuming to be practical. However, in recent years, autonomous learning methods have gained renewed attention because the overall performance of hardware has improved, including increased memory and CPU performance, and it has become very easy to collect, store, and share data anywhere through mobile connectivity, including cloud computers.
Amidst these changes, a branch of self-learning called ‘deep learning’ is leading the way. Deep learning is a learning method that has a similar structure to the human brain. The human brain has neurons connecting each part of the brain, and it is known that as the brain learns, the connections between the parts that are used more often and related to each other become stronger. This is the inspiration for deep learning, which, like neurons, has a connection structure between each piece of learning, and this structure is constantly changing as additional training data comes in, creating feedback between the pieces of learning. Therefore, it is possible to recycle previous learning and adapt to new environments based on it.
The present and future of machine learning
Both autonomous and supervised learning methods are currently being used in different contexts. Supervised learning is used for fast development that only needs to be applied in a specific environment, while autonomous learning is used in fields that require machines that can be applied in a variety of environments. However, one thing is clear: there is a lot of research going on in both directions.
Recently, many global companies, including Facebook and Google, have been investing a lot of energy in the development of machine learning. Facebook is working hard to improve its facial recognition machine using supervised learning, and Google recently acquired a venture led by deep learning experts called DeepMind for about 440 billion won. The fact that the company was only three years old and hadn’t sold a single product shows how much Google is investing in machine learning. Currently, machine learning is only capable of recognizing objects, so it’s a long way from being able to have conversations with humans with emotions, like the operating system in the movie Her. However, the rapid pace of recent advancements suggests that the day of talking to Her is not far off.
The possibilities for machine learning are endless. In healthcare, AI systems are being developed to help diagnose and treat patients, and in finance, systems are being introduced to predict market fluctuations and make investment decisions. The technologies that make our lives easier, such as self-driving cars, smart homes, and voice assistants, are also becoming more sophisticated and intelligent as machine learning advances. We are living in an era where machine learning is constantly evolving, and its future advancements are unimaginable.
Therefore, machine learning is not just a technology that processes and analyzes data, but is expected to enrich human life and contribute to solving various problems. Machines that can emotionally connect with humans, like in the movie “Her,” may seem like something from the distant future, but the possibility is becoming more and more real. Machine learning is already deeply embedded in our daily lives, and future developments will bring us more surprises and innovations.
