Deep learning is one of the technologies used in the AI field. It is a machine learning method using a multi-layered neural network . By securing a sufficient amount of data and allowing it to learn, AI can automatically extract features from the data.
Deep learning has made it possible to learn unstructured data (images, natural language, sound), which was difficult to digitize in the past, and can be used for image recognition, voice recognition, natural language processing, and anomaly detection . rice field.
This article provides an easy-to-understand explanation of the meaning and mechanism of deep learning, typical algorithms, how to use it, and practical examples.
What is deep learning?
I will explain the meaning and mechanism of the term deep learning.
Meaning of deep learning
Deep learning is a learning method that utilizes neural networks, a typical machine learning algorithm.
★ What is a neural network?
Neural networks are AI modeled on neurons (nerve cells in the brain)
A neural network is an AI that models the structure and function of neurons (the nerve cells that make up the brain of living organisms).
A neural network consists of an input layer that receives data, an intermediate layer (hidden layer) that processes the weights flowing from the input layer, and an output layer that outputs the results.
How deep learning works
Deep learning is one of neural network technologies
In deep learning, the neural network itself can automatically extract the features of the data group as long as there is enough training data.
A multi-scale hidden layer cuts the input data into various sizes and identifies features, so based on the given data, it extracts detailed patterns, large structures, and overall contours.
It excels at pattern recognition of data that cannot be encoded, such as images .
Difference from machine learning
Machine learning is “learning” in AI. It has the meaning of “the machine itself learns” like a human learns.
Machine learning is one of the technologies that support AI, and deep learning is one of the machine learning methods .
Representative algorithms of deep learning
Among the methods of deep learning, representative ones are introduced.
CNN (Convolutional Neural Network)
CNN (Convolutional Neural Network) is mainly used for image recognition and motion detection. It consists of a “convolution layer” that extracts image features and a “pooling layer” that analyzes the features. It has a high pattern recognition ability for images and is characterized by being able to identify quickly.
RNN (Recurrent Neural Network)
RNN (recurrent neural network) is a general term for neural networks with an autoregressive structure, and can handle variable length time series data. It is mainly used for speech recognition, video recognition, and natural language processing.
How to use deep learning
Deep learning has made it possible for AI to learn unstructured data (images, natural language, sounds) that were previously difficult to digitize .
The increased variety of digitalization has improved the accuracy of optimization and recommendations. Currently, it is used in various situations such as image recognition and voice recognition.
The main uses of deep learning are as follows.
Usage (1) Image recognition
Technology that recognizes human faces and characters from images. It separates the features from the input image or video background and extracts the features of the target object.
Examples: iPhone face recognition, Facebook tagging
Usage (2) Speech Recognition
Technology that recognizes human voices. You can identify people by voice input or voice.
Example: Siri, Alexa voice input
Usage (3) Natural language processing
Technology that allows computers to understand written and spoken language used in everyday communication.
Examples: machine translation, language modeling, answering questions
Usage ④ Anomaly detection
Technology that detects anomalies using time-series data collected from sensors.
Examples: credit card fraud, quality control in manufacturing
Now, the technology of attention “GAN (hostile generation network)”
Generative Adversarial Networks (GANs) are attracting particular attention among technologies that utilize deep learning .
A GAN is a kind of generative model and consists of two networks: a generative network and a discriminative network . A major feature is that the generation side learns to deceive the identification side, and the identification side learns to identify more accurately. The name “adversarial” is used because the two networks learn with opposing goals.
It is possible to generate non-existent data when creating an image, or to convert existing data according to its characteristics. For example, if you use a tool that utilizes GAN called “IMAGE INPAINTING” published by NVIDIA, you can erase people and objects from the image and process only the background.
I was able to erase the person who should have existed and create a background that should not have existed.
For example, if the purpose is to generate an image, the generation side outputs the image, and the identification side determines whether it is correct or not. The generator learns to deceive the discriminator, and the discriminator learns to discriminate more accurately. In this way, the two networks are called adversarial because they learn with conflicting goals.
By using deep learning, it is possible to move with natural facial expressions just by reading a picture of a person.
Example of work efficiency improvement using AI image recognition (Unimate)
Automatic measurement app that utilizes AI image recognition
Unimate Co., Ltd. is a company that develops rental uniform business. The company had a problem with frequent erroneous orders due to measurement errors.
Monster Lab has started developing an automatic measurement application that utilizes AI image recognition to solve the problem . We have developed an original AI engine by deriving a method of “creating a 3D model from an image and predicting the actual size from it” from technical research.
We repeated verification using the measurement data provided by Unimate, improved the accuracy of AI image recognition, and succeeded in reducing costs caused by human errors such as incorrect measurements .
Summary: Deep learning enables learning on unstructured data
Deep learning is a machine learning method that uses multilayered neural networks .
Deep learning has made it possible to learn unstructured data such as images and voices that could not be digitized in the past, making it possible for machines to perform image recognition, voice recognition, natural language processing, and anomaly detection. These technologies are used in various services and products.
