AI, Machine Learning, and Deep Learning: What Can They Do?

First, let’s understand how these terms are related.
Put simply, the relationship can be described as “AI > Machine Learning > Deep Learning.”
Machine learning is a subfield of AI, and deep learning is one of the many techniques within machine learning.

AI stands for “Artificial Intelligence” and is a conceptual term. It generally refers to machines that are designed to replicate human intelligence and cognitive abilities.
Its interpretation varies among researchers and institutions, and there is no single, universally accepted definition.

AI can be broadly categorized into “Narrow AI” and “General AI,” or alternatively, “Weak AI” and “Strong AI.”

In contrast, “General AI” refers to AI that is not limited to specific tasks or domains, but can handle a wide range of problems at a level equal to or beyond human capabilities.
Humans are able to respond to unexpected situations by making comprehensive judgments based on past experiences and solving problems accordingly.
At present, the methods for realizing General AI have not yet been clearly established. However, in Japan, organizations such as Artificial General Intelligence Society are actively working on its development.

The classification of “Strong AI” and “Weak AI” was proposed by American philosopher John Searle in his 1980 paper “MINDS, BRAINS, AND PROGRAMS”. This distinction is based on whether AI possesses human-like consciousness and intelligence.
“Strong AI” refers to AI that has self-awareness like humans and can learn, make decisions, and process information with its own intent. In contrast, “Weak AI” does not possess self-awareness and therefore cannot respond to unexpected situations that have not been pre-programmed.

The difference between “Strong AI” and “General AI” appears to lie in whether the artificial intelligence has self-awareness.

Machine learning refers to “a mechanism that enables machines to discover patterns and rules from data on their own.”
Among the types of AI mentioned earlier, it is categorized as a form of Narrow AI.
There are several types of learning methods within machine learning.

“Supervised learning” is a method in which the model is trained using data that has been labeled with correct answers.

Typical problems addressed by supervised learning include “regression,” which predicts continuous numerical values, and “classification,” which predicts the class to which a given data point belongs.

An example of “regression” is learning the relationship between data such as average temperature or weather conditions and product sales, and using it to predict future sales quantities.

An example of “classification” is spam email filtering.
By learning from labeled data that distinguishes between spam and non-spam, the system identifies patterns in the text and predicts whether new incoming emails are spam.

At Weathernews, in addition to the traditional AMeDAS data used for weather forecasting, various independently collected datasets are used in supervised learning models to improve forecast accuracy.

Efforts to Improve Forecast Accuracy - Weathernews

“Unsupervised learning” is a method in which the model is trained without providing correct answers in the training data.
For example, when a large volume of emails is processed using unsupervised learning, they can be grouped (clustered) based on similarities in textual features.

Kewpie utilizes AI on its food production lines to detect defective raw materials. By applying unsupervised learning, the system is trained only on normal (non-defective) products, and anything outside of that is identified as an anomaly.

Detecting Defective Products on Food Production Lines Using Artificial Intelligence (BrainPad Inc.)

“Reinforcement learning” is a method in which a machine learns optimal strategies through trial and error by actually taking actions. It is particularly suited to tasks such as games or gambling, where outcomes take time to appear or require many repeated attempts.

“AlphaGo,” which became the first AI to defeat a professional human Go player without a handicap and sparked a global AI boom, is a system based on reinforcement learning.

AlphaGo Official Website

“Deep learning” refers to a machine learning method that utilizes “deep neural networks,” which are composed of multiple layers of “neural networks”—models inspired by the human brain’s neural circuitry.

In traditional machine learning, humans need to determine and adjust which elements of the data influence the outcome. In contrast, with deep learning, given a sufficient amount of data, machines can automatically extract features from the data without human intervention.
As a result, systems can learn patterns that humans may not be able to identify, leading to significant advancements in fields where human perception and judgment previously had limitations.

So, in what kinds of fields is deep learning actually applied?
Here are a few examples of its practical uses.

Image recognition is a technology that inputs images or videos to identify and detect features such as text or faces. It works by extracting distinguishing features from the background, performing matching and transformations, and ultimately identifying the target features.

Autonomous driving technology in automobiles uses image recognition to analyze road conditions.
In Japan, as of April 2020, the sale of vehicles equipped with Level 3 autonomous driving—where the system handles driving under limited environments or traffic conditions—was officially permitted.
Honda’s “Legend” became the world’s first vehicle to receive approval for Level 3 autonomous driving technology in November 2020, with plans for commercial release within the same fiscal year.

Automated Drive | Honda

Speech recognition is a technology that enables the identification and processing of spoken language. It can recognize human speech and convert it into text, as well as analyze vocal characteristics to identify the speaker.

Empath analyzes the physical features of voice data to detect emotions in real time, independent of language.
By integrating this technology into call center systems, it can visualize the emotions of both customers and operators. It can also be embedded in robots to enable more natural communication between humans and machines.

Empath Official Website

Natural language processing (NLP) is a technology that enables computers to process and understand the natural language humans use in everyday communication, including both written and spoken language.

Alexa, which is built into devices such as Amazon Echo, combines NLP with speech recognition to analyze spoken input, understand its meaning, and perform a wide range of tasks accordingly.

What You Can Do with Alexa | Amazon

As the name suggests, prediction refers to technology that uses artificial intelligence to forecast future events. This is made possible by analyzing vast amounts of historical data.

SPAIA analyzes data from sports such as baseball and soccer to predict game outcomes, player movements, and strategies.

SPAIA | Sports × AI × Data Analytics Media

Going forward, as AI technologies continue to advance, opportunities for their application will only increase.
In addition, with the expansion of 5G networks, the types and volume of data that can be collected will grow, further broadening the potential uses of AI.

In this context, it is important to develop a clear understanding of each of these technologies. We hope this article serves as a helpful guide in that process.