Since March 2016, interest in AlphaGo and artificial intelligence has skyrocketed, and advances in artificial intelligence technology, including the concepts of deep learning and machine learning, have been in the spotlight. Deep learning is based on artificial neural networks to process data, and Deep Mind’s Deep Q Network is a prime example of its use. With the development of AI comes the need to think about social and ethical issues.

In March 2016, unless you live in the middle of nowhere, cut off from the rest of the world, you’ve probably heard about AlphaGo and artificial intelligence at least once. You’ve probably looked up the terms deep learning, machine learning, and machine learning, wondering why they’re so famous. But how many of us know even a little bit about them? Most of us know the names of a few words like big data, but not much more than that. In this article, we’ll take a look at the history and technical theoretical underpinnings of “deep learning,” one of the hottest topics in artificial intelligence right now, and talk briefly about the “Deep Q Network,” the brainchild of the deep minds behind AlphaGo.

Overview and history of deep learning

Before we talk about deep learning, let’s talk about “learning”: machine learning. According to American computer scientist Tom M. Mitchell, machine learning can be defined as “the use of experience with a task to improve the performance or outcome of that task”. Getting better with repeated experience is what machine learning is all about. Just as humans get better with training, machine learning is designed to mimic how humans “learn”.
Deep learning is a type of machine learning based on artificial neural network technology. An artificial neural network is a model in the field of artificial intelligence that mimics a neural network consisting of a large number of neurons in the brain, each of which sends and receives signals through synapses, and deep learning is designed with multiple layers in the hierarchical structure of an artificial neural network. These layers are what makes deep learning different from traditional machine learning. While traditional machine learning requires the extraction of basic features in advance to learn from the input data for the desired outcome, deep learning can work with very basic data (such as pixels in the case of images) to produce an output without extracting features.

The evolution of deep learning and its modern importance

The concept of deep learning, or artificial neural networks, was discussed more than 30 years ago. It is usually considered to have started with Frank Rosenblatt’s Perceptron in the mid-1950s, which was a single-layer neural network that worked for linear models but not for nonlinear ones, and the hardware of the time could not keep up with the computations. However, it was later shown that deep neural networks with multiple layers, such as deep learning, could analyze nonlinear models and solve problems such as overtraining, which led to a resurgence of interest.
In the modern era, deep learning has become one of the fastest-growing fields, with significant advances in hardware and big data. This has led to innovations in a variety of applications, including self-driving cars, speech recognition, image recognition, and more. For example, self-driving cars use deep learning to analyze road conditions in real-time to determine safe driving routes. Speech recognition technology is also using deep learning to more accurately recognize and understand human speech. These technologies are transforming our daily lives, and their impact is expected to continue to grow.

Deep Mind’s Deep Q Network

Deep Mind, better known to most of us as AlphaGo, uses an algorithm called the Deep Q Network (DQN) in its artificial intelligence. DQN is a combination of deep neural networks and reinforcement learning. The theory of reinforcement learning is similar to utility theory in economics, which states that the rational behavior is to calculate the utility of each object and choose the one with the highest utility value. However, when behavior changes as a result of learning, it means that utility values change. Reinforcement learning theory is a theory that clarifies how utility values change with experience. It just changes the terminology from utility to a value function.
The beauty of DQN is that it can perform with very simple information, and it performs well in most competitive or challenging tasks. Deep Mind demonstrated this by applying it to classic games on the Atari 2600: in the case of a block-breaking game, the only information given was the score and the screen. At first, the machine was not very good at the game, but it started to get better and better scores and eventually learned to get better scores on its own.

The future of deep learning and our role

So far, we’ve covered the history of deep learning, its theoretical foundations, and Deep Mind’s Deep Q Network. The fact that it’s easier to feed data into than any other machine learning in the past and still produce great results is what makes deep learning so exciting right now. Some people see in deep learning the potential for “strong artificial intelligence,” which is a step up from the current state of artificial intelligence and surpasses humans at everything. In fact, it has already surpassed humans in some areas.
The era of AI is not far away. At a time like this, it’s more important than ever to learn about AI and ask yourself what you should be doing in this era. With AI impacting our daily lives and industries across the board, we need to keep up with the advances in technology, but also think deeply about the ethical and social issues it raises. For example, there are concerns that advances in AI could lead to job losses. The challenge for us will be to figure out how to deal with this and find ways to coexist with AI.