What is the difference between deep learning and machine learning? – Also explains the positioning in AI!

The term “deep learning ” is used more and more often.

The development of AI has become a social focus, and AI is currently experiencing its third boom, but deep learning is a technology that is particularly attracting attention.

But how is deep learning different from conventional technology and how can it be used? This time, I will explain the features of deep learning when compared to conventional technology.

Difference between machine learning and deep learning

What is machine learning

Machine learning is a technology that allows a machine to learn its rules by itself by learning a huge amount of data, enabling advanced predictions and judgments.

There are three types of learning methods: supervised learning that learns a large amount of data and automatically acquires its features, unsupervised learning that classifies data in various dimensions, and reinforcement learning that obtains the correct answer by repeating trial and error. .

Positioning of machine learning in AI

Machine learning is a branch of AI

Machine learning is one of the elemental technologies of AI.

AI is a general term for machines with intelligence that can think, judge, and act like humans, and machine learning is one of the means to realize it.

What is deep learning

Deep learning is a machine learning technology that enables flexible decision-making according to the situation by learning a large amount of data and automatically extracting common feature values.

It is characterized by being able to analyze with higher accuracy than conventional machine learning.

Deep learning is modeled after human brain cells

Deep learning is modeled after human brain cells (neurons).

First, as a structure, the input layer, hidden layer, and output layer are each composed of nodes, and the nodes of each layer are connected by threads called edges. Information enters from the input layer, travels along edges, passes through hidden layers, and exits from the output layer.

In addition, each node has a weight value set, and the passing information is multiplied by that value and passed to the next node.

In deep learning , the weight values ​​are optimized to give the most accurate answer.

Positioning of deep learning in AI

Deep learning is a part of supervised learning within machine learning.

Machine learning that uses neural network technology is called deep learning.

Advantages of deep learning over machine learning

Features can be automatically extracted

Unlike conventional technologies, deep learning is characterized by the ability to automatically extract feature quantities. Its accuracy is also high, so deep learning can find features that are invisible to the human eye.

In conventional technology, humans taught AI feature values

In the conventional technology, when learning data, a human being instructed the feature value to be focused on.

For example, when creating an AI that distinguishes images of cats and dogs, humans were set in advance to pay attention to the shape of the ears and faces.

However, the shape of the ears and the shape of the face cannot be uniformly determined. In some cases, it was necessary to set a large amount of feature values, and there was a limit to human power alone.

However, deep learning enables automatic extraction of feature values, which dramatically improves accuracy.

What Machine Learning Can Do and Use Cases

(1) Regression

Regression is a type of supervised learning in machine learning, and is a learning method for handling continuous numerical values ​​such as “sales” and “growth rate”.

For example, past customer data can be used to predict how many times a new customer will visit a store, and past data can be used to predict future numbers.

Sales and demand forecast | DeNA

DeNA has introduced a function to the taxi dispatch app “MOV” that predicts the supply and demand of taxis with AI and suggests routes to places where there are likely to be many users.

②Classification

Classification is a method of determining what category, class, or type the data to be analyzed belongs to, and is one of the ” supervised learning ” in machine learning.

Classification is used as a learning method to classify and predict information about labels such as “dog” and “cat” and categories such as “buy” and “do not buy”.

Spam/fraud detection | NEC

NEC has started providing the cloud service “AI Transaction Screening Support Service” that supports screening of unfair trading. This service was adopted by SBI SECURITIES.

(3) Dimensional reduction

In machine learning as well, the so-called “curse of dimensionality” occurs when there are too many features, resulting in poor accuracy. Dimensionality reduction is a technique for reducing the number of dimensions (number of features) of data.

Data Visualization｜Trial

TRIAL has installed 1,500 AI cameras independently developed by its subsidiary Retail AI at its store in Shingu, Fukuoka Prefecture, in order to analyze the purchasing behavior of visitors and use it for sales promotion. In the future, we aim to expand annual sales from 6 billion yen to 10 billion yen.

④Clustering

Clustering is an extension of classification, and is a typical unsupervised learning method that collects similar data groups by function or category.

Text mining tool | Nittele, DoCoMo

Nippon Television and DoCoMo plan to develop an automatic summarization system for news articles using AI and have on-site staff conduct performance evaluations.

⑤ Recommendation

Recommendation literally means “recommendation”.

Various related behavioral data and attribute data are accumulated on the Web. This data may be entered directly by the customer, such as member information, or may be behavior-tracked, such as browsing information or purchased product information.

Product Recommendation | Netflix

The recommendation algorithm is said to use data such as movie genres, categories, search keywords, work ratings, and viewing time.

Netflix’s basic algorithm is to input this complex member behavior data, personalize it with machine learning, and provide it in the best way for each viewer.

Netflix’s basic algorithm does not include information such as gender or age. When you sign up for Netflix, the first step is to listen to some of your favorite movies and collect your preferences. This information is used for personalization, but the most recent data is actually more important.

(6) Anomaly detection

Anomaly detection is a technology for detecting and estimating machine failures and abnormal values ​​in data analysis. The use of techniques such as data mining to identify observations or expected patterns that are inconsistent with the rest of the data in a dataset.

Anomaly detection uses machine learning to find outliers that do not occur under normal conditions.

Anomaly/failure detection | NTT DoCoMo, Automagi

NTT Docomo and Automagi have developed a system that detects when a driver is asleep.

When it detects that you are falling asleep, it will alert the driver with a buzzer from your smartphone. This system does not require a dedicated terminal and can be used simply by installing a dedicated application on a smartphone.

What deep learning can do and examples of its use

image recognition

Image recognition is a technology that can determine what kind of image even an unknown image is by learning a large amount of image data with deep learning .

Its utilization is not limited to images on the web, but it is also active in every part of the real world, such as cameras.

[Case study] Detecting diseases with deep learning! Imaging diagnosis of Fujifilm

Fujifilm is developing a medical system that automatically detects affected areas using image recognition.

In the past, many diseases in the body were discovered by doctors, but the human eye was inevitably overlooked or misidentified.

However, by utilizing AI image recognition, it will be possible to detect affected areas with greater accuracy than humans.

voice recognition

Speech recognition is a technology in which AI analyzes speech data and extracts features as speech.

For example, Siri installed in Google Home and iPhone uses speech recognition technology that listens to human speech and understands the meaning with text AI.

[Usage example] AI automatically transcribes speech! Smart Secretary

Transcription work, which involves listening to people’s voices and converting them into characters, inevitably takes time and effort.

However, if you use Smart Writing, you can save the trouble of transcribing because it will transcribe the voice you heard as it is.

In business, there are many parts that can be used, such as creating minutes.

natural language processing

Natural language processing is a technology that allows machines to understand the natural language that we use as a communication tool on a daily basis. Conventionally, computers could only understand programming languages, but by making AI learn the patterns and meanings of natural languages, it will be possible to communicate naturally with humans.

[Usage example] Gmail auto-reply function

Gmail has an automatic reply function that makes use of natural language processing.

AI understands the contents of the received mail and automatically generates some simple reply sentences accordingly.

You don’t have to think about the contents of the reply from the beginning, so it will lead to more efficient email operations.

predict

AI-based prediction is a technology that predicts values ​​and results that can be realized in the future based on data.

AI demand forecasting is already being used in fields such as restaurant demand forecasting and transportation congestion forecasting, enabling more effective purchasing and measures than forecasts that rely on human intuition and experience.

[Usage example] Demand forecast for conveyor belt sushi restaurant Sushiro

Conveyor-belt sushi chain “Sushiro” uses AI demand forecasting to predict sales after 1 minute and 15 minutes.

By accumulating big data on the sales situation collected from all stores and taking into account factors such as store congestion and customer seating times, we have achieved a high degree of accuracy.

By using AI, it becomes possible to analyze big data in real time, and it is also applied to reducing food loss and marketing and product planning.

Cases where development using deep learning is suitable

Deep learning enables complex analysis that humans cannot judge.

For that purpose, we need a huge amount of training data and hardware that can withstand computational processing.

If you can prepare a large amount of data, you can also consider building an AI model.

As an advantage of deep learning, I mentioned that a computer can extract features instead of humans. increase.

For example, in the case of life-threatening work such as the use of AI in medical care, it is important to be able to explain why the AI ​​made such a diagnosis or judgment.

From this point of view, it is necessary to decide whether to use deep learning or not.

summary

Deep learning refers to some techniques of machine learning.

Deep learning, which enables conventional machine learning to learn more autonomously, is used in all areas.

And in the future, deep learning will be used more than ever, and I think it will improve the convenience of our lives.