The success of the artificial intelligence program AlphaGo has put neural network learning in the spotlight. Neural network learning is a technique that models the human brain to learn from data and use it to solve complex problems. It’s opening up the possibility that artificial intelligence can think like humans.
In 2016, an event occurred that shocked many people. Alpha Go, an artificial intelligence Go program developed by Google’s DeepMind, defeated South Korean professional Go player Lee Sedol in a five-game match, winning 4 out of 5 games. This event sparked a surge of interest in artificial intelligence across society. At the same time, concerns about the possibility of a dystopian world where machines dominate humans, which has been the subject of many novels and movies, began to grow.
AlphaGo is a program based on the theory of machine learning, a branch of computer science. Machine learning is the idea that programs can learn to perform specific tasks without relying on explicit programming. This approach can be necessary in a variety of fields. For example, tasks such as categorizing documents into topics, predicting stock market trends, or recognizing numbers on a license plate are not always easy to clearly identify and formalize the sequence of steps that must be taken to achieve the goal.
Over the years, various methods based on machine learning theory have been devised and utilized. Scientists are constantly trying to utilize machine learning theory to develop artificial intelligence that is at least as intelligent as humans. There have been several attempts, but so far, they have not been very successful. However, recent progress has been dramatic, as evidenced by the AlphaGo example. The idea behind this is a concept called neural network learning.
To understand neural network learning, we first need to look at the structure of the human brain. This is because neural network learning is an approach that attempts to solve problems by modeling the human brain.
The human brain is made up of tiny units called neurons. There are about 100 billion neurons in the human brain, and they are intricately connected to form a network. Each neuron is responsible for receiving and processing information from other neurons, and sending that information back to other neurons.
Neurons are composed of dendrites, cell bodies, and axons. Dendrites are the part that receives information from neighboring neurons. They’re shaped like branches, so they can receive information from multiple neurons at the same time. The cell body is responsible for integrating and processing this information. The axon is a thin tube that transmits the processed information from the cell body to neighboring neurons. The information transfer process that takes place within a nerve cell is all electrical.
Neurons interact with neighboring neurons without direct contact, but with some space between them, which is called a synapse. At the synapse, information is transmitted chemically via neurotransmitters. When an electrical signal reaches the end of an axon in a nerve cell, neurotransmitters are released into the synapse. Neighboring nerve cells are then stimulated by the neurotransmitters, and an electrical signal is transmitted from the tip of the dendrite to the axon. In order to stimulate neighboring neurons, the amount of neurotransmitter released must be above a certain level.
Despite being composed of such relatively simple structural units, the human brain is capable of complex thinking and feeling a wide range of emotions. Programs that use neural network learning have abstract neural networks that reflect the essential characteristics of biological neural networks. The abstract neural networks in computer programs are called artificial neural networks to distinguish them from biological neural networks.
Like biological neural networks, artificial neural networks are made up of neurons. Each neuron can receive multiple inputs from other neurons, perform certain computations based on these inputs, and output the result as an output. The computational process of a neuron consists of two steps.
The first step is to multiply each input by a weight and simply add them together. When you look at the connections between neurons in a biological neural network, some connections are stronger than others, and some are weaker. In an artificial neural network, weights reflect these differences in the strength of the connections between neurons. This way, we can distinguish between information that is relatively important and information that is not, and influence the computational results differently.
The second step is to apply an activation function to the values obtained to derive new values. The activation function is based on a certain value and gives a value close to 1 if the input value is greater than that value, and close to 0 otherwise. The value that the activation function gives depends on the influence it has on the next neuron. To use an analogy from a biological neural network, the activation function is responsible for determining the amount of neurotransmitter release.
These simple neurons are organized into layers, which are then stacked on top of each other to form an artificial neural network. For example, consider the problem of recognizing an animal given a picture of it. The computer first converts the image file of the animal into a set of numbers, which it then feeds into the neural network. The neural network then performs a series of calculations to determine whether the animal is more likely to be a cat, a dog, or something else. The results of the computation are highly dependent on the weights defined by the connections between the neurons. The performance of a neural network is determined by how well the weights are set. The process of optimizing the weights using data is called neural network training. Of course, we use the word “learning” because this process is performed by a program, not a human.
The concept of neural network learning has been around for a long time, so why has it only recently gained so much attention? There are two reasons. First, programs based on neural network learning require a significant amount of computation, depending on the context. In the past, it was difficult for computers to perform enough computations to produce meaningful results, but thanks to technological advances, this is now possible. Second, the development of the internet and mobile devices has led to the accumulation of a wide variety of data in various forms, including text, images, voice, and video, which has made it easier to obtain the large amounts of data needed to train neural networks. As a result, neural network learning is rapidly emerging as a central topic in academia.
Until now, when creating a program to achieve a specific goal, humans had to explicitly define every step of the way, but with neural network learning, this need has been reduced to some extent, and problems that were difficult to solve with traditional methods can now be solved. Scientists are excited about the promise of neural network learning and believe that it could eventually be applied to develop artificial intelligence that thinks in a similar way to humans.
