[Reading Science] AI, Can It Work Without Big Data and Deep Learning?

How does AI distinguish between '먹다' (to eat) and '만찬' (dinner)? Until now, it has differentiated them through extensive data and deep learning. In the future, however, even without prior learning experience, AI will be able to transform and interpret already stored data to analyze the characteristics between problems it has learned and those it has not, thereby distinguishing them.
[Photo by YouTube screen capture]

[Asia Economy Reporter Kim Jong-hwa] It has been several years since artificial intelligence (AI) began to dominate our daily lives. The fact that AI defeated a professional Go player is considered a milestone in human history. This growth of AI can be attributed to the long-term learning of data accumulated through countless hours and costs.

AI can be defined as a computer system created by humans that possesses intelligence similar to that of humans. AI analyzes patterns and rules contained in data through algorithms within computer programs by learning, and infers the outcomes of actions. The representative methods of creating AI are machine learning and deep learning.

Machine learning is a method where humans provide rules and answers for each problem to train the AI, while deep learning uses artificial neural networks similar to the human brain to enable the machine to learn on its own. However, recently, AI that does not require such learning processes has emerged and attracted attention.

What is the biggest difference between humans and AI? For example, when identifying that a Sonata is a sedan among cars, AI needs pre-existing data categorized by category to recognize the object. It must learn or search big data about what a car is, how it looks, and what types exist.

On the other hand, humans do not necessarily need to search data. With just a few pieces of information, humans intuitively judge that a Sonata is a sedan. Of course, basic knowledge about cars is required in this case, but it does not require extensive information data like AI.

This special AI learning ability inspired by ordinary human capabilities is called "zero-shot learning." Through zero-shot learning, AI can infer correct answers with very little information like humans, without the arduous time of big data and deep learning.

This means AI development can be achieved through unsupervised learning without large amounts of data, computing resources, or human intervention. A method surpassing deep learning, which is the mainstream of AI development, has been developed. Although deep learning shows excellent performance in natural language processing (NLP) and image classification, it also faces many economic challenges.

It is not easy to secure big data sufficient for machine learning in all industrial fields, and even if secured, enormous costs arise in refining and processing the data to make it usable, leaving various challenges. Zero-shot learning is a new method that can achieve effects comparable to deep learning with a small amount of data.

Zero-shot learning is an advanced method developed from "transfer learning," widely used in machine learning, which learns to find correct answers by utilizing relationships and commonalities between data.

For example, if the answer to problem A also exists in other problems B, C, and D, AI analyzes the relationships among problems A to D to identify commonalities. Then, by analyzing these commonalities, it finds the correct answer. This allows AI to infer answers even when data is not abundant.

The place that most appropriately utilizes this method is Google's "Google Neural Machine Translation (GNMT)." The system infers translations between languages it has never translated before by using translation data between other languages. Even without prior learning experience, it transforms and interprets stored data to analyze features between learned and unlearned problems and predicts answers.

The era of judging AI performance by the amount of learned data is coming to an end. If the data learning process in AI development is greatly reduced, tremendous time and costs can be saved. Will AI that surpasses human capabilities appear soon?