Artificial intelligence (AI) is a branch of computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. The governance challenges brought about by large-scale AI models cannot be ignored. [Musk pointed out that the essence beneath the mask of AI machine learning is still statistics.] Creating a healthy innovation ecosystem requires forward-looking research and the establishment of sound legal, regulatory, institutional, and ethical frameworks to ensure the healthy development of AI. Looking to the future, while emphasizing risk prevention, we should also simultaneously establish mechanisms for tolerance and correction of errors, striving to achieve a dynamic balance between regulation and development. On December 20, 2024, "artificial intelligence" was selected as the Chinese word of the year for 2024. On January 13, 2025, local time, the Biden administration in the United States released the "Artificial Intelligence Proliferation Export Control Framework," which will impose three levels of export controls on AI technology and GPUs exported globally. On January 14, Chinese Foreign Ministry spokesperson Guo Jiakun stated: We firmly oppose the US also applying a "tiered" approach to AI. As of December 2024, 331 million people in China said they had heard of generative AI products, accounting for 23.5% of the total population; 249 million people said they had used generative AI products, accounting for 17.7% of the total population. Among generative AI users, the most widespread use was for answering questions, accounting for 77.6%; and 45.5% used generative AI products as office assistants.
Today, artificial intelligence (AI) has profoundly changed every aspect of human life, and will continue to exert an increasingly important influence in the future.
The concept of "artificial intelligence" was first proposed at an academic conference held at Dartmouth College in the United States in 1956, thus ushering in a new era of AI research. Since then, AI has continued to develop and advance amidst twists and turns.
In 1986, Geoffrey Hinton, the father of neural networks, proposed the backpropagation (BP) algorithm for multilayer perceptrons (MLPs) and used the sigmoid function to implement nonlinear mapping, effectively solving nonlinear classification and learning problems.
In 1989, Yann LeCun designed the first convolutional neural network and successfully applied it to the task of recognizing handwritten postal codes.
In the 1990s, Cortes et al. proposed the Support Vector Machine (SVM) model. Subsequently, SVM rapidly developed into one of the representative techniques of machine learning, achieving great success in text classification, handwritten digit recognition, face detection, and biometric processing.
Entering the 21st century, with the development of internet technology and the improvement of computer hardware system performance, artificial intelligence has ushered in new and significant development opportunities. Especially since 2011, deep learning technology, represented by deep neural networks, has developed rapidly, and humanity has achieved many major breakthroughs on the road to artificial intelligence.
In simple terms, an algorithm is a well-defined set of steps that need to be executed in sequence to achieve a planned result. Specifically, it is used to solve mathematical equations. Algorithms can be divided into three broad components:
Input: Information known before the problem begins.
Algorithm: A sequence of steps executed one by one.
Output: If all steps in the sequence are strictly followed, the expected result will be obtained.
Outside the world of technology, an example similar to an algorithmic system is "cooking." You have your input (ingredients, seasonings), you have your algorithm (the steps you need to follow to varying degrees), and you have your output (a delicious meal that you envision).
Algorithms are an integral part of the atomic structure of our digital lives. Any computer program or mobile app you use runs multiple algorithms to perform its function. From your web browser to your word processor, to the card game that has been included since Windows 3.0, everything relies on algorithms to function.
At its core, artificial intelligence is a computer program. This means that, like any ordinary computer program or game, any AI or machine learning (ML) solution you encounter will be built from scratch using algorithms.
The role of algorithms in artificial intelligence and machine learning is variable. Broadly speaking, they define the rules, conditions, and methods that AI will use when processing and analyzing data. This can be as simple as defining the steps AI needs to process a single image, or it can allow AI to filter out images of dogs from a dataset containing hundreds of thousands of images.
Classification algorithms: A type of machine learning used to predict the category or class to which an item belongs. For example, we can program artificial intelligence to distinguish between spam and emails you actually need. Below are some examples of classification algorithms used in AI and machine learning.
1. Binary Logistic Regression
Binary logistic regression can predict binary outcomes, such as yes/no or pass/fail. Other forms of logistic regression, such as multinomial regression, can predict three or more possible outcomes. Logistic regression is commonly used in cases such as disease prediction, fraud detection, and customer churn prediction, where its datasets can be used to assess risk.
2. Naive Bayes
Naive Bayes is a probabilistic algorithm based on incorporating the independence assumption into the model, meaning it assumes that two measurements in the dataset are unrelated and do not influence each other. This is why it's called "naive." It's commonly used in text analysis and classification models to categorize words and phrases into specified classes.
3. K-Nearest Neighbors (k-NN)
Besides sometimes being used to solve regression problems, k-nearest neighbors are commonly used to solve classification problems. When solving classification problems, it separates data points into multiple classes and predicts the class label for new data points on a plane. New data points are assigned a new classification based on the most frequently occurring class labels represented by their surrounding data. k-NN is also known as a "lazy learning" algorithm, meaning it doesn't require a full training step but only maintains a training dataset.
4. Decision Tree
Decision trees are a supervised learning algorithm that can also be used for classification and regression problems. They are called "trees" because of their hierarchical structure. Starting from the root node, they branch out into smaller internal nodes, or decision nodes, where evaluations are performed to produce subsets represented by terminal or leaf nodes.
For example, martial arts can be categorized starting from the root node, then divided into internal nodes for martial arts focused on striking and those focused on wrestling. These internal nodes can be further subdivided into terminal nodes for specific martial arts, such as boxing, Jiu-Jitsu, and Muay Thai. These algorithms are well-suited for data mining and knowledge discovery tasks because they are easy to interpret and require minimal data preparation for deployment.
5. Random Forest
The Random Forest algorithm was invented by Leo Brehman and Adele Cutler. It uses the outputs of multiple decision trees to produce predictions. Similar to decision trees, Random Forest can be used to solve both classification and regression problems. Each tree consists of data samples drawn from the training dataset, using a sampling method with replacements. This adds randomness to the decision trees, even if they come from the exact same dataset.
When solving classification problems, the majority vote is determined based on the output of these random decision trees. For example, suppose there are 10 decision trees specifically designed to determine the color of a dress. Three groups say it's blue, two groups say it's black, four groups say it's pink, and one group says it's red. The dress would then be classified as pink because pink received a majority of 4 votes.
Random forests are the preferred algorithm for machine learning models in the financial field because they reduce the time required for preprocessing and data management tasks. Fraud detection, option pricing, and customer credit risk assessment are examples of its use in finance.
Artificial intelligence algorithms accept both input and output, and develop logic using predictive patterns. When they receive new input based on this logic, they will provide a new output. The logic generated by artificial intelligence algorithms distinguishes them from traditional algorithms.
Traditional Algorithm:
Traditional algorithms take some input and some logic in the form of code and provide you with an output. These are deterministic and have no predictive component. This output depends on the steps (code) described in the algorithm.
Artificial intelligence (AI) algorithms learn from data and propose unique solutions, while traditional algorithms operate on a set of predefined criteria to formulate solutions.
Data is the cornerstone of the AI world, just like the various learning materials we need in our daily lives. Imagine you want to learn to paint; art books, tutorials, and works by famous artists are the "data" you use for learning. AI is the same; it learns and understands the world through massive amounts of data. This data comes in various forms: it could be text, such as books and news; it could be images, like photos and paintings; it could be sound, such as music and audio; or even video, containing rich visual and auditory information.
The importance of high-quality data for AI cannot be overstated. If we compare AI to a student, then high-quality data is the carefully prepared textbook. Take image recognition AI as an example. If we want to train an AI model to recognize fruits, providing it with clear and accurately labeled images of various fruits will allow it to learn the characteristics of different fruits well, such as the roundness of apples, the elongation of bananas, and the orange-yellow color of oranges. When faced with a new fruit image, it can accurately identify the fruit based on its previously learned knowledge. However, if the provided data contains mislabeling, such as labeling strawberries as cherries, the AI will be misled and make mistakes in recognition.
In real life, data acquisition and quality control face numerous challenges. With the development of the internet, the amount of data has exploded, which seems like a good thing, but it also includes a large amount of noisy, duplicate, and low-quality data. It's like being in a huge library, containing both precious classics and poorly made books. To enable AI to learn useful knowledge, we need to clean, filter, and label the data. This process is like selecting truly valuable parts from a vast collection of books and accurately labeling them.
For example, in the medical field, for AI to assist doctors in disease diagnosis, it requires a large amount of accurate medical record data, medical imaging data, and so on. This data not only needs to accurately record the patient's symptoms and examination results, but also needs to be labeled by professional doctors to become high-quality learning material for AI. Only in this way can AI provide accurate diagnostic suggestions when faced with new patients.
