Understanding Artificial Intelligence, Machine Learning, and Deep Learning

When it comes to the popular terms of recent years, “Artificial Intelligence” must be on the list. Following the explosive popularity of ChatGPT last year, “AI (Artificial Intelligence)” has repeatedly topped the trending search lists and has been named the Word of the Year 2023 by Collins, a British dictionary publisher.

In addition to “Artificial Intelligence”, we often hear about “Machine Learning” and “Deep Learning”…

What do these terms mean? What is their relationship? Let’s follow along to learn more~~

Artificial Intelligence – Artificial Intelligence

When it comes to artificial intelligence, people’s first reaction might be the robots with human-like intelligence from science fiction movies, but in reality, artificial intelligence is not just about robots.

Artificial Intelligence was proposed by John McCarthy in 1956, and the initial definition was “the science and engineering of making intelligent machines”.

Today’s artificial intelligence refers to “a new technology science that studies and develops theories, methods, techniques, and application systems used to simulate, extend, and enhance human intelligence”.

It sounds a bit convoluted, doesn’t it? Let me summarize: artificial intelligence is about enabling machines to simulate human thinking abilities, allowing machines to perceive, think, and even make decisions like humans.

Today, artificial intelligence is no longer a mere discipline but an interdisciplinary field involving computer science, psychology, linguistics, logic, philosophy, and more.

Artificial intelligence may seem like a profound technology, but it is actually a broad concept. Around us, there are already various forms of artificial intelligence, such as autonomous driving, facial recognition, smart robots, machine translation, and so on.

When faced with various forms of artificial intelligence, we can classify them into three categories based on their capabilities:

• Weak Artificial Intelligence (Artificial Narrow Intelligence, ANI)

  AI that excels in a specific area and can only perform specific tasks. For example, a facial recognition system can only recognize images; if you ask it about the weather tomorrow, it won’t know how to respond.

• Strong Artificial Intelligence (Artificial General Intelligence, AGI)

  AI that is similar to human-level intelligence, capable of exhibiting intelligence similar to humans across multiple domains, understanding, learning, and executing various tasks. Currently, strong AI has not yet been realized and remains a long-term goal of AI research.

• Super Artificial Intelligence (Artificial Superintelligence, ASI)

  AI that surpasses human intelligence, capable of outperforming humans in various fields and executing any intellectual task. Although super AI frequently appears in science fiction, in reality, it is just a theoretical concept and has not yet been realized.

At this point, I would like to ask everyone: what category does AlphaGo, which defeated the world champion in Go, belong to?

Machine Learning – Machine Learning

As mentioned earlier, the goal of artificial intelligence is to enable machines to think and make decisions like humans. How is this achieved?

Think back to when we were born; we knew almost nothing. After decades of learning, we acquired various knowledge and skills.

Machines are the same; to enable them to think, they must first learn, summarizing patterns from experience to develop a certain level of decision-making and discrimination ability. This is the core of artificial intelligence – machine learning.

Machine learning specifically studies how computers can simulate or achieve human learning behaviors, acquiring new knowledge and skills through learning, thereby reorganizing existing knowledge structures and continuously improving their performance.

Machine learning is a multidisciplinary field that involves probability theory, statistics, approximation theory, algorithm complexity theory, and other disciplines.

How do machines learn? Let’s first look at the human learning process:

1. Class: Learning theoretical knowledge and inputting knowledge

2. Summarizing and Reviewing: Reinforcing understanding through review

3. Organizing Knowledge Framework: Structuring knowledge into a system

4. Homework: Deepening understanding through practice

5. Weekly Tests: Checking mastery

6. Filling Gaps: Improving learning methods

7. Final Exam: Checking final learning outcomes

The learning process of machines is also similar, consisting of the following seven steps:

1. Data Acquisition: Collecting relevant data

2. Data Processing: Transforming data and standardizing data formats

3. Model Selection: Choosing the appropriate algorithm

4. Model Training: Training the model using data and optimizing the algorithm

5. Model Evaluation: Evaluating model performance based on prediction results

6. Model Adjustment: Adjusting model parameters to optimize performance

7. Model Prediction: Making predictions on unknown result data

In short, machine learning is about automatically inducing logic or rules from data through algorithms and using the induced results to make predictions with new data.

For example, if we want a computer to identify a dog when it sees one, we need to show the computer a large number of pictures of dogs and tell it that this is a dog.

After extensive training, the computer will summarize certain patterns and when it sees a dog next time, it captures the corresponding features and concludes, “This is a dog”.

If the algorithm is not perfect, it might mistakenly identify a cat as a dog, which requires the computer to improve the algorithm automatically through experience data, thus enhancing its predictive ability.

Based on the learning method, machine learning can be divided into the following four categories:

• Supervised Learning

  Learning from labeled data, where the data contains independent and dependent variables, making predictions based on known input and output data, such as classification tasks and regression tasks.

  Classification Tasks: Predicting the category of data, such as spam detection, identifying species of plants and animals, etc.

  Regression Tasks: Predicting data based on previously observed data, such as predicting housing prices, height, and weight, etc.

• Unsupervised Learning

  Analyzing unlabeled data, where the data contains only independent variables without dependent variables, discovering patterns in the data, such as clustering and dimensionality reduction.

  Clustering: Grouping similar items together without focusing on what those items are, such as customer segmentation.

  Dimensionality Reduction: Compressing high-dimensional data into a lower-dimensional representation by extracting features, such as combining the mileage and years of use of a car into a wear value.

• Semi-Supervised Learning

  Training data has only partial labels, first using unsupervised learning to process the data, then using supervised learning to train and predict the model.
  For example, a phone can recognize a person’s photo (unsupervised learning), and when that person’s photo is labeled, subsequent new photos of that person will also be automatically labeled (supervised learning).
• Reinforcement Learning

  Optimizing algorithms through interaction with the environment based on rewards or penalties until the maximum reward is achieved, resulting in optimal strategies. For example, a vacuum robot optimizes its cleaning path after hitting an obstacle.

Deep Learning – Deep Learning

Through the above understanding, I believe everyone is now familiar with machine learning. So what about deep learning? What is its relationship with machine learning?

Deep learning is a new research direction in the field of machine learning, an algorithm that learns and understands complex data through multi-layer neural networks.

Machines learn complex tasks by learning the deep representations of sample data, ultimately gaining analysis and learning abilities similar to humans, capable of recognizing text, images, and sounds.

Unlike traditional machine learning, deep learning uses neural network structures, where the length of the neural network is referred to as the model’s “depth”; therefore, learning based on neural networks is called “deep learning”.

Neural networks simulate the neural networks of the human brain, where neuron nodes can process and transform data. Through multi-layer neural networks, the features of data can be continuously extracted and abstracted, allowing machines to better solve various problems.

The typical deep learning algorithms include the following four types:

• Convolutional Neural Network (CNN): Often used for image recognition and classification tasks.
• Recurrent Neural Network (RNN): Suitable for processing sequential data, such as natural language processing.
• Long Short-Term Memory (LSTM): A special RNN structure that can better handle long sequence data.
• Generative Adversarial Network (GAN): Used for generating new data, such as images, audio, or text.

With the support of deep learning, artificial intelligence is developing rapidly, and I believe that in the near future, we will enter a new AI era.

Conclusion

Useful knowledge has increased again; let me summarize briefly:

• “Artificial Intelligence” is a broad concept aimed at enabling machines to think and perform tasks like humans.
• “Machine Learning” is a method to achieve artificial intelligence, aimed at learning patterns from data, with traditional machine learning requiring manual determination of data features.
• “Deep Learning” is a specific branch of machine learning based on neural networks, capable of automatically learning data features.

I believe that through today’s learning, everyone will no longer be confused~~

In addition to the three terms introduced today, there are many other AI-related terms; what else do you know? Feel free to discuss in the comments~~

WeChat ID｜ITP-CAS

Open, Integrate, Seek Truth, Innovate

· Chinese Academy of Sciences·

· Institute of Theoretical Physics·