In this blog post, we’ll explore the basic concepts of deep learning and examine exactly how it works, using the example of training a computer to recognize letters written in various handwriting styles.
‘Will Doctors Become an Extinct Profession?’, ‘Humans Are No Longer Needed’, ‘The Arrival of Learning Machines’, ‘The Future of Jobs Transformed by Robots and Artificial Intelligence’… As these book titles suggest, there was a great deal of interest at the time in the advancement of artificial intelligence and the resulting changes to human life and the job market. After AlphaGo defeated 9-dan Lee Sedol 4–1 in their 2016 match, interest surged even further in technologies that implement human learning, reasoning, perception, and natural language understanding capabilities through computer programs. This interest led to the question of how the development of artificial intelligence would impact humanity’s future. So, what exactly is artificial intelligence, and how does it affect people’s daily lives? To answer these questions, it is necessary to understand deep learning, one of the primary methods used to build artificial intelligence. In this article, after providing a basic introduction to deep learning, we will examine how it works in practice using the example of training a computer to recognize letters of the alphabet.
First, deep learning refers to a field of artificial intelligence technology that enables computers to learn from data and identify patterns in a manner similar to humans. For example, when it comes to the 26 letters of the alphabet from A to Z, humans have the ability to recognize them as the same letter even if their shapes differ slightly. However, it is not easy for computers to recognize letters written in various handwriting styles as the same letter as naturally as humans do. Deep learning is a technology that can be used to overcome these limitations. It is a machine learning technology based on artificial neural networks that enables computers to learn patterns on their own using large amounts of data. By mimicking the human brain’s information-processing method—which involves discovering patterns within vast amounts of information and then distinguishing objects—computers are trained to discern data in a similar manner. By applying deep learning technology, computers can identify necessary features and patterns based on training data and make inferences about new inputs without humans having to manually define criteria for determining what an object is. Because of these characteristics, deep learning is already being utilized in various fields, such as image and speech recognition and photo analysis.
In summary, deep learning is a machine learning technology that enables computers to derive appropriate results for new inputs by learning the relationships between specific input values and their corresponding outputs. Using the alphabet mentioned earlier as an example, let’s assume we are designing a computer capable of accurately recognizing letters written in any handwriting style. The goal here is to take an image of a random letter as input and obtain 26 output values—each representing the probability that the image corresponds to one of the 26 letters from A to Z.
The specific method can be explained as follows. First, the computer is fed data consisting of thousands to tens of thousands of letter images. However, rather than simply inputting the images, correct answer information—indicating which letter each image represents—is also provided. This process, up to this point, is the preprocessing stage, which organizes the data and prepares it for training. Next, multiple layers are formed to ensure that a given letter image ultimately reaches the desired letter. Since these layers mimic the structure of the human brain, where countless neurons are interconnected to transmit signals, they are also called artificial neural networks. The number and size of these layers vary depending on the range and size of the given input and output values, as well as the complexity of the problem. Each layer can be viewed as a structure that applies weights and biases to the input data and performs various mathematical operations; as the image passes through each layer, its features are extracted step by step. In the initial layers, relatively simple features such as lines or contours are identified, while deeper layers combine these features to learn more complex shapes and patterns. Ultimately, after passing through all layers, the model provides a result indicating the probability that the given image represents a specific letter of the alphabet. Since the weights required for training are not properly set in the initial layers, it is difficult to accurately derive meaningful results. Once training begins, numerous training images pass through multiple layers to reach the final output—that is, the probability values corresponding to each letter. The system then calculates the difference between the predicted result and the actual answer, and uses a backpropagation method to determine how that error was influenced by the weights of each layer, thereby adjusting the weights incrementally. As this process is repeated, the operations performed by each layer evolve into values better suited for analyzing and distinguishing letters, ultimately increasing the likelihood of accurately predicting the correct letter for a given input image.
Artificial neural networks in deep learning can consist of an input layer, multiple hidden layers, and an output layer. This structure can be applied to the alphabet recognition example mentioned earlier. In this case, an alphabet image composed of 32×32 pixels is used as the input; as the image passes through multiple layers, its features are gradually extracted, and the final classification layer is configured to output N = 26 probability values. In actual image recognition, structures such as convolutional neural networks (CNNs) are widely used; by processing the pixel information of an image across multiple layers, features such as lines, contours, and textures are extracted step by step, ultimately enabling the image to be classified.
In conclusion, deep learning is a technology that trains computers to recognize and classify data in a manner similar to humans. Deep learning is of great significance because it can learn from vast amounts of data—including images, audio, and facial data—and identify patterns within them. Furthermore, since it is a technology applicable to various fields closely related to human life, the general public also needs to understand the basic principles of deep learning and embrace it with an open mind. Currently, deep learning is being utilized in various fields, such as automatic speech recognition, human face recognition, consumer-related marketing methodologies, and automated recommendation systems, and its scope of application is expanding further into areas such as computer vision, speech recognition, natural language processing, generative AI, and robotics. In the past, the possibility of deep learning creating machines that surpass humans was often viewed as a distant future scenario, but today, there are already instances where artificial intelligence exceeds human capabilities in specific domains. A prime example is AlphaGo’s victories against the world’s top Go players, and systems that emerged afterward, such as AlphaZero, have demonstrated performance that surpasses the highest human levels not only in Go but also in chess and shogi. If research into deep learning continues steadily, we are likely to see more advanced artificial intelligence and robots that complement human capabilities in various fields and even surpass humans in specific areas.
