There was a time when Japan was developing artificial intelligence at the risk of its national fortune. At that time, AI was a 'rule-based system' that attempted to imitate the logical thinking process of humans. In terms of technology, it focused on knowledge representation and logical reasoning. In short, it was a method of teaching computers a lot of 'if ~, then ~' rules and using those rules to find the correct answer. AI systems process knowledge based on explicit rules and logic.
This is called 'deductive reasoning' or 'top-down approach' and is a typical rule-based methodology. It may be effective in fields where the goal is 'automation by robots', but it is difficult to introduce in cases where there are no 'drawing or expressing pictures that you like' or 'clear rules and goal settings that apply to everyone'. Specifically, it means that tasks such as 'creating beautiful images using various expressions' or 'creating in collaboration with humans' are difficult to implement because they lack accuracy and clarity.
Modern AI is not 'rule-based', but is mainly built on machine learning and deep learning technologies. This technology can be called a 'bottom-up approach' or 'inductive reasoning' that learns patterns from large amounts of data and makes predictions, inferences, and decisions. Thanks to the application of new technologies, AI can outperform humans in tasks where conditions are diverse and goal setting is difficult, such as image recognition, natural language processing, and game play. If you show a computer a lot of information and data, the computer can acquire a model through its specialty, 'iterative machine learning', and predict new things or solve problems.
In 2024, Google's Transformer was introduced into ChatGPT and Stable Diffusion, which are widely used as general AI services and AI models. This is based on numerous books and sentences, and learns patterns from general facts about humans, animals, and events, and establishes models to infer them. The environment has clearly changed from AI in the past in terms of massive computational power, data, and the relevance between applications and society.
In the rule-based era, large-scale computing resources could not be used at the speed they are now. In addition, before the advent of the Internet, large-scale data sets for training AI systems were also limited. Therefore, abstract problem-solving methods that did not require complex calculations were preferred. Meanwhile, the current development of AI was achieved through computational power that has significantly increased thanks to GPUs and massive data sets secured through the Internet. Through these developments, complex and diverse AI models could be trained, and unknown tasks that could not be solved in the past could be challenged.
In terms of applications, from language translation, which was originally a specialty of AI, to professional systems for law enforcement or medical fields, to surveillance camera analysis, autonomous driving, real-time multilingual translation, advanced speech recognition, and personalized recommendation systems, numerous applications closely related to daily life are being developed. In particular, tasks such as image or document generation were previously considered 'worthless' or 'too difficult to set goals', but now large-scale inference models that can surpass human creativity have appeared in the field of document and image generation. ChatGPT and Stable Diffusion are representative examples. These are now called AI or generative AI, and will be defined as entities that further expand human activities starting from 2024.
Then, let's study the flow of technological development from machine learning utilizing large-capacity computational resources and large-scale data sets that form the basis of these technologies, to basic knowledge related to artificial neural networks (NN), which are the basis of these technologies, to cutting-edge 'deep learning (DL)'.
The Relationship Between Changing Society and AI |
In addition to the advancement of AI, the 'relationship between society and AI' is also undergoing a major change due to the emergence of generative AI. There are fierce debates all over the world about whether to include machine learning and inference-based generation in the existing copyright concept, whether to include one's own works in the responsibility and learning target, and how to pay royalties (compensation). Stable Diffusion has also gone through a process of adapting to and reaching agreements on social issues as it has evolved from its initial version to the latest generation, and has undergone numerous updates. Each country is enacting laws related to artificial intelligence. |
