I. What Technologies Does Artificial Intelligence Actually Include?
Artificial Intelligence (AI) is not a single algorithm or technology, but a comprehensive discipline covering multiple research directions and technical fields.
From early rule-based reasoning and expert systems, to machine learning and deep learning, and then to today's large language models, generative AI, and agents, artificial intelligence has undergone multiple evolutions of technical paradigms.
Current mainstream AI technologies include:
- Symbolic AI
- Machine Learning (ML)
- Deep Learning (DL)
- Generative AI
- Natural Language Processing (NLP)
- Computer Vision (CV)
- Speech Technology (Speech AI)
- Reinforcement Learning (RL)
- Knowledge Representation and Reasoning
- Multimodal AI
- Embodied AI and Robotics
- AI Agent Technology
One concept needs to be clarified first:
These technologies do not belong to the same classification dimension.
For example:
- Machine learning describes how models learn from data.
- Deep learning describes a class of learning methods based on multi-layer neural networks.
- Generative AI describes models' ability to generate new content and the corresponding technical direction.
- Natural language processing describes the task domain in which artificial intelligence processes human language.
- Multimodal AI describes a model's ability to process and relate information across multiple modalities.
- Agent describes a system form that performs planning, decision-making, and tool execution around a goal.
Therefore, these technologies cannot simply be regarded as peer-level, unrelated classifications.
II. The Overall Technical System of Artificial Intelligence
From the dimensions of technical paradigms, learning methods, model capabilities, and application domains, a more easily understandable AI technical system can be established.
1. Overall Classification
| Classification Dimension | Main Technologies | Problems Solved |
|---|---|---|
| AI Technical Paradigms | Symbolism, Connectionism, Behaviorism | How to build intelligent systems |
| Learning Methods | Supervised Learning, Unsupervised Learning, Self-Supervised Learning, Reinforcement Learning | How to learn from data and interaction |
| Model Technologies | Machine Learning, Deep Learning, Neural Networks, Transformer | How to build and train models |
| Content Generation | LLM, Diffusion, GAN, VAE | How to generate text, images, audio, and video |
| Perception and Understanding | NLP, CV, Speech Recognition | How to understand external information |
| Reasoning and Decision-Making | Knowledge Reasoning, Planning, Search, Reinforcement Learning | How to analyze problems and choose actions |
| System Form | Multimodal AI, Agent, Multi-Agent | How to organize model capabilities to complete complex tasks |
| Interaction and Execution | Robotics, Embodied AI, Tool Calling | How to interact with physical or digital environments |
2. Technical System Relationship Diagram
flowchart TB
A["Artificial Intelligence (AI)"]
A --> B["Technical Paradigms"]
A --> C["Learning Methods"]
A --> D["Models and Algorithms"]
A --> E["Capabilities and Application Domains"]
A --> F["Intelligent Systems"]
B --> B1["Symbolism"]
B --> B2["Connectionism"]
B --> B3["Behaviorism"]
C --> C1["Supervised Learning"]
C --> C2["Unsupervised Learning"]
C --> C3["Self-Supervised Learning"]
C --> C4["Reinforcement Learning"]
D --> D1["Traditional Machine Learning"]
D --> D2["Deep Learning"]
D --> D3["Generative Models"]
E --> E1["Natural Language Processing"]
E --> E2["Computer Vision"]
E --> E3["Speech Technology"]
E --> E4["Multimodal AI"]
F --> F1["Large Language Models"]
F --> F2["AI Agent"]
F --> F3["Embodied AI"]
classDef root fill:#DCEFE5,stroke:#4F8B70,color:#24563F,stroke-width:2px
classDef group fill:#EAF3EC,stroke:#83A995,color:#315D49
classDef item fill:#F5F8F4,stroke:#B5CBBB,color:#40594A
class A root
class B,C,D,E,F group
class B1,B2,B3,C1,C2,C3,C4,D1,D2,D3,E1,E2,E3,E4,F1,F2,F3 item
This diagram provides an overview of the knowledge system and does not imply strict containment relationships among the various technologies. For example, large language models are a type of deep learning model, typically pretrained using self-supervised learning, and can also be post-trained in combination with reinforcement learning.
The following starts from the core technologies and gradually works through their principles and relationships layer by layer.
III. The Three Major Technical Paradigms of Artificial Intelligence
Looking at the history of AI development and its technical ideas, classic AI research can usually be understood through three paradigms:
- Symbolism
- Connectionism
- Behaviorism
They represent different approaches to building intelligent systems.
1. Symbolism: Achieving Intelligence Through Rules and Logic
Symbolism holds that intelligence can be expressed and reasoned about through symbols, knowledge, and logical rules.
Its core idea is:
Represent knowledge about the real world as symbols, then derive new conclusions through logical rules.
For example, a simple rule-based system:
Rule 1: All people age
Rule 2: Zhang San is a person
Inference result: Zhang San will age
The system does not need to learn this conclusion by training on large amounts of data; instead, it reasons directly based on predefined knowledge and rules.
Main technologies include:
- Logical Reasoning
- Rule-based System
- Expert System
- Knowledge Representation
- Automated Planning
- Search Algorithms
Typical applications:
- Medical expert systems
- Rule engines
- Automated theorem proving
- Business decision systems
- Knowledge reasoning systems
Technical characteristics:
Advantages: - The reasoning process is relatively explicit. - Rules and knowledge are highly interpretable. - Suitable for domains with clear rules and well-defined constraints.
Limitations: - Relies on manually constructed knowledge and rules. - Difficult to handle complex, vague, and uncertain information. - In open environments, rule maintenance costs are high.
Symbolism has not disappeared in modern AI. Technologies such as knowledge graphs, rule engines, constraint solving, and neuro-symbolic AI still draw heavily on this idea.
2. Connectionism: Learning Intelligence Through Neural Networks
Connectionism holds that intelligence can emerge from the connections and coordination among large numbers of simple computational units.
Its key ideas derive from an abstraction of biological nervous systems.
Artificial Neural Networks (ANNs) are an important technical foundation of connectionism.
A simple neuron can be abstracted as:
where:
- x_i: input data.
- w_i: the corresponding weight.
- b: bias.
- f: activation function.
- y: output result.
By forming networks from large numbers of neurons and using training algorithms to adjust network parameters, a model can gradually learn the patterns in the data.
Main technologies include:
- Artificial Neural Network (ANN)
- Convolutional Neural Network (CNN)
- Recurrent Neural Network (RNN)
- Transformer
- Deep Neural Network (DNN)
Typical applications:
- Image recognition
- Speech recognition
- Natural language processing
- Large language models
- Image and video generation
Connectionism is an important foundation of current deep learning and large model technologies.
Its advantage is that it can automatically learn complex patterns from large amounts of data, without needing humans to write rules for every situation.
However, its internal decision-making process is usually difficult to explain directly, and it also places high demands on data, computing power, and training methods.
3. Behaviorism: Learning Behavior Through Environmental Interaction
Behaviorism emphasizes the interaction between an agent and its environment, as well as the continuous adjustment of strategies based on behavioral outcomes.
The question it focuses on is not how to explicitly represent all knowledge, but rather:
Given an environment and a goal, how should actions be chosen to obtain better results?
Reinforcement Learning (RL) is an important technical representative of this idea.
For example, training a robot to complete a walking task:
- The robot performs actions.
- The environment produces feedback based on the actions.
- The system receives rewards or punishments.
- The agent adjusts its strategy.
- Task performance improves through continuous interaction.
Main technologies include:
- Markov Decision Process (MDP)
- Q-Learning
- Deep Q-Network (DQN)
- Policy Gradient
- Actor-Critic
- Proximal Policy Optimization (PPO)
Typical applications:
- Game AI
- Robot control
- Autonomous driving decision-making
- Resource scheduling
- Large model post-training
It should be noted that behaviorism as an AI technical paradigm is related to the behaviorist ideas in psychology, but the two should not be simply equated.
4. Comparison of the Three Paradigms
| Comparison dimension | Symbolism | Connectionism | Behaviorism |
|---|---|---|---|
| Core idea | Symbols and logical reasoning | Neural network learning | Environment interaction and behavior optimization |
| Knowledge source | Human-defined rules and knowledge | Data training | Interaction feedback and rewards |
| Core technology | Logic, rules, search | Neural networks, deep learning | Reinforcement learning, policy optimization |
| Main advantages | Interpretable, clear rules | Strong expressive power, adapts to complex data | Can optimize sequential decisions |
| Main limitations | High cost of knowledge construction | Weak interpretability | High training cost and exploration difficulty |
Real-world AI systems do not necessarily adopt only one paradigm.
For example, a modern Agent can use neural networks to understand user requests, use search and planning algorithms to organize tasks, and then use rules to constrain the execution process.
Such systems that combine different paradigms also reflect the convergence trend in artificial intelligence technology.
IV. Machine Learning: Letting Computers Learn from Data
Machine learning (ML) is one of the most important technical fields in artificial intelligence.
Traditional programs usually have processing rules explicitly written by developers:
Input data
↓
Execute predefined rules
↓
Output result
Machine learning takes a different approach:
Through training data and learning algorithms, the model automatically learns the patterns between inputs and outputs.
1. Basic Principles of Machine Learning
Machine learning usually contains three core elements:
- Data: samples used for training and evaluation.
- Model: a mathematical structure used to represent patterns in data.
- Learning algorithm: a method for adjusting model parameters based on data.
Take house price prediction as an example:
Given input features such as house area, location, and floor, the model learns the relationship between these features and house prices from historical housing transaction data, and ultimately predicts the price of a new house.
It can be abstracted as:
Where:
- x: input features.
- f_\theta: model with parameters \theta.
- \hat{y}: model prediction result.
The training process usually requires defining a loss function:
Where:
- y: true label.
- L: a loss function that measures the difference between predicted results and true results.
- \theta^*: optimized model parameters.
The core task of machine learning is to use training algorithms to find suitable model parameters so that the model can maintain good performance on unseen data.
2. Main Categories of Machine Learning
Based on the training data and feedback method, machine learning is usually divided into the following categories.
2.1 Supervised Learning
Training data contains inputs and corresponding target labels.
The model completes prediction tasks by learning the mapping relationship between inputs and labels.
Main tasks:
- Classification
- Regression
Typical algorithms:
- Linear Regression
- Logistic Regression
- Decision Tree
- Random Forest
- Support Vector Machine (SVM)
- Gradient Boosting Trees (GBDT, XGBoost)
For example:
Use data already labeled as spam and normal email to train a model, so that the model can identify whether new emails are spam.
2.2 Unsupervised Learning
Training data has no explicit target labels.
The model attempts to discover patterns from the structure of the data itself.
Main tasks:
- Clustering
- Dimensionality Reduction
- Density Estimation
- Anomaly Detection
Typical algorithms:
- K-Means
- DBSCAN
- PCA
- Gaussian Mixture Model
For example:
Based on users' purchasing behavior, automatically divide users into different groups without needing to predefine which category each user belongs to.
2.3 Semi-supervised Learning
Train using both a small amount of labeled data and a large amount of unlabeled data at the same time.
The main idea is to use structural information in unlabeled data to improve the model's learning performance.
Typical methods:
- Pseudo Labeling
- Consistency Regularization
- Label Propagation
Suitable for scenarios where labeling costs are high and unlabeled data is abundant.
2.4 Self-supervised Learning
Self-supervised learning constructs training objectives from the data itself and does not require manual labeling of each item.
For example:
In language model training, part of the text is used as a prediction target, and the model predicts the hidden or subsequent content based on context.
Common training tasks:
- Masked Language Modeling
- Next Token Prediction
- Contrastive Learning
- Image Masked Modeling
Self-supervised learning is an important foundation of modern large model pretraining.
2.5 Reinforcement Learning
Reinforcement learning learns through continuous interaction between an agent and its environment.
Core elements include:
- Agent: the intelligent agent.
- Environment: the environment.
- State: the state.
- Action: the action.
- Reward: the reward.
- Policy: the policy.
Its goal is usually to maximize long-term cumulative reward.
Where:
- \pi: policy.
- r_t: reward obtained at time t.
- \gamma: discount factor.
- T: time horizon.
Reinforcement learning is widely used in areas such as robot control, game decision-making, and resource scheduling.
3. Overview of Machine Learning Methods
flowchart TB
A["Machine Learning (ML)"]
A --> B["Supervised Learning"]
A --> C["Unsupervised Learning"]
A --> D["Semi-supervised Learning"]
A --> E["Self-supervised Learning"]
A --> F["Reinforcement Learning"]
B --> B1["Classification"]
B --> B2["Regression"]
C --> C1["Clustering"]
C --> C2["Dimensionality Reduction"]
C --> C3["Anomaly Detection"]
D --> D1["Small amount of labeled data"]
D --> D2["Large amount of unlabeled data"]
E --> E1["Predictive training objectives"]
E --> E2["Contrastive Learning"]
F --> F1["Environment interaction"]
F --> F2["Reward feedback"]
F --> F3["Policy optimization"]
classDef root fill:#DCEFE5,stroke:#4F8B70,color:#24563F,stroke-width:2px
classDef group fill:#EAF3EC,stroke:#83A995,color:#315D49
classDef item fill:#F5F8F4,stroke:#B5CBBB,color:#40594A
class A root
class B,C,D,E,F group
class B1,B2,C1,C2,C3,D1,D2,E1,E2,F1,F2,F3 item
4. A Commonly Confused Concept: Self-Supervised Learning Is Not the Same as Unsupervised Learning
The two are often discussed together, but they are not exactly the same.
Unsupervised learning usually does not rely on explicit human labels; its focus is on discovering data structure.
Self-supervised learning, by contrast, constructs supervisory signals from the data itself to form explicit training objectives.
For example:
- K-Means performs clustering based on data features and is an unsupervised learning method.
- BERT is trained by predicting masked words and is a self-supervised learning method.
- GPT is pretrained by predicting the next token and is also a self-supervised learning method.
In terms of training form, self-supervised learning has supervisory signals, but these signals are not obtained through manual item-by-item annotation.
This is a very important distinction for understanding how modern large models are trained.
V. Deep Learning: The Core Technical Foundation of Modern AI
Deep Learning (DL) is an important branch of machine learning.
It learns complex features and representations in data through multi-layer neural networks.
Traditional machine learning often requires manually designed features, whereas deep learning can learn feature representations layer by layer from raw data through multi-layer networks.
1. Core Principles of Deep Learning
A typical neural network consists of multiple layers:
- Input Layer
- Hidden Layer
- Output Layer
Data passes through multiple computational layers in the network, gradually completing feature transformations.
flowchart LR
A["Input Data"] --> B["Input Layer"]
B --> C["Hidden Layer 1"]
C --> D["Hidden Layer 2"]
D --> E["Hidden Layer 3"]
E --> F["Output Layer"]
F --> G["Prediction Result"]
style A fill:#DCEFE5,stroke:#4F8B70,color:#24563F
style B fill:#EAF3EC,stroke:#83A995,color:#315D49
style C fill:#EAF3EC,stroke:#83A995,color:#315D49
style D fill:#EAF3EC,stroke:#83A995,color:#315D49
style E fill:#EAF3EC,stroke:#83A995,color:#315D49
style F fill:#EAF3EC,stroke:#83A995,color:#315D49
style G fill:#DCEFE5,stroke:#4F8B70,color:#24563F
During training, the network computes the loss based on the prediction result and the target result, then calculates parameter gradients through backpropagation (Backpropagation), and updates the model parameters using an optimization algorithm.
A typical training process can be summarized as follows:
flowchart TB
A["Prepare Training Data"] --> B["Forward Propagation"]
B --> C["Compute Loss"]
C --> D["Backpropagation"]
D --> E["Compute Parameter Gradients"]
E --> F["Optimizer Updates Parameters"]
F --> G{"Training Condition Met?"}
G -->|"No"| B
G -->|"Yes"| H["Obtain Trained Model"]
style A fill:#DCEFE5,stroke:#4F8B70,color:#24563F
style B fill:#EAF3EC,stroke:#83A995,color:#315D49
style C fill:#EAF3EC,stroke:#83A995,color:#315D49
style D fill:#EAF3EC,stroke:#83A995,color:#315D49
style E fill:#EAF3EC,stroke:#83A995,color:#315D49
style F fill:#EAF3EC,stroke:#83A995,color:#315D49
style G fill:#F7F4E8,stroke:#C7B77A,color:#675A2D
style H fill:#E5F0E9,stroke:#83A995,color:#315D49
Two concepts need to be distinguished:
- Forward propagation is responsible for computing the model output.
- Backpropagation is responsible for computing the gradients of the loss with respect to the model parameters.
Backpropagation itself does not directly update parameters; parameter updates are usually performed by optimizers such as SGD and Adam.
2. Main Network Architectures in Deep Learning
2.1 CNN: Convolutional Neural Network
CNN (Convolutional Neural Network) is mainly used to extract features from data with spatial structure.
Core components:
- Convolutional layer (Convolution Layer)
- Activation function (Activation Function)
- Pooling layer (Pooling Layer)
- Fully connected layer (Fully Connected Layer)
Typical applications:
- Image classification
- Object detection
- Image segmentation
- Medical image analysis
An important characteristic of CNN is that it reduces the number of parameters through local receptive fields and weight sharing, while extracting local spatial features.
2.2 RNN: Recurrent Neural Network
RNN (Recurrent Neural Network) is used to process data with sequential relationships.
Through a recurrent structure, it passes the hidden state from the previous time step to the current time step, enabling the model to use historical information.
Typical applications:
- Time series forecasting
- Early speech recognition
- Text sequence modeling
Traditional RNNs are prone to vanishing or exploding gradients when processing long sequences.
LSTM and GRU are important improved architectures proposed to address the long-term dependency problem of traditional RNNs.
2.3 Transformer: Attention Mechanism Architecture
Transformer is an important foundation of modern large-model technology.
It is centered on the self-attention mechanism (Self-Attention), enabling the model to compute relationships between different positions in a sequence.
Compared with traditional RNNs, Transformer is better suited to large-scale parallel training.
Core components include:
- Token Embedding
- Positional Encoding or positional representation mechanism
- Multi-Head Attention
- Feed Forward Network
- Residual Connection
- Layer Normalization
Transformer has become an important architectural foundation for modern large language models, multimodal models, and many generative models.
2.4 GNN: Graph Neural Network
GNN (Graph Neural Network) is used to process graph-structured data.
Graph data usually consists of nodes (Node) and edges (Edge).
For example:
- Social network relationships.
- Knowledge graphs.
- Molecular structures.
- User–item relationships in recommendation systems.
The core idea of GNN is to learn representations of nodes, edges, or entire graphs through information passing and aggregation between nodes.
3. Comparison of Different Deep Learning Architectures
| Architecture | Core Characteristics | Typical Applications |
|---|---|---|
| CNN | Local receptive field, weight sharing | Image recognition |
| RNN | Recurrent state propagation | Sequence modeling |
| LSTM / GRU | Gating mechanisms, long-term dependency modeling | Time series, speech |
| Transformer | Self-attention, parallel training | Large language models, multimodal |
| GNN | Graph-structured information aggregation | Knowledge graphs, recommendation |
It should be noted that network structures and application domains do not correspond one-to-one.
For example, Transformer is not only used for natural language processing, but is also widely used in vision, speech, and multimodal tasks.
VI. Generative Artificial Intelligence: From Recognition and Prediction to Content Generation
Generative AI is an important direction in the current development of AI technology.
Traditional discriminative models usually focus on:
- Which category does the input belong to?
- Does the image contain a certain object?
- Is this transaction anomalous behavior?
Generative models, by contrast, focus on:
- How to generate a piece of text?
- How to generate an image?
- How to generate a piece of audio?
- How to generate a video?
The two solve different problems, but they are not completely opposed technologies.
1. What Is Generative AI?
Generative AI refers to artificial intelligence technology that can learn the distribution or structure of training data and generate new content based on input conditions.
Generated objects can include:
- Text
- Images
- Audio
- Video
- Code
- 3D content
Its core is not simply copying training data, but learning the statistical patterns, structures, and latent representations in the data, and generating new samples accordingly.
2. Main Technical Routes of Generative AI
2.1 Autoregressive Model
Autoregressive models predict the next Token or data element step by step based on the existing context.
Taking large language models as an example:
Where:
- x_t: The Token currently being generated.
- x_{<t}: The context before the current Token.
- P(x_t|x_{<t}): The probability of predicting the next Token based on the existing context.
The model gradually generates the complete text by continuously predicting the next Token.
Typical applications:
- Large language models.
- Code generation.
- Text continuation.
- Dialogue systems.
The GPT series models are important representatives of autoregressive language models.
2.2 GAN: Generative Adversarial Network
GAN (Generative Adversarial Network) consists of two core networks:
- Generator: responsible for generating samples.
- Discriminator: responsible for determining whether a sample is real data or generated data.
The two networks are continuously optimized through adversarial training.
flowchart TB
A["随机噪声"] --> B["生成器 Generator"]
B --> C["生成样本"]
D["真实样本"] --> E["判别器 Discriminator"]
C --> E
E --> F["判别结果"]
F --> G["对抗训练与参数优化"]
G --> B
G --> E
classDef start fill:#DCEFE5,stroke:#4F8B70,color:#24563F
classDef model fill:#EAF3EC,stroke:#83A995,color:#315D49
classDef process fill:#F5F8F4,stroke:#B5CBBB,color:#40594A
class A,D start
class B,E model
class C,F,G process
GAN was once widely used in tasks such as image generation, image enhancement, and style transfer.
Main difficulties include training instability, mode collapse, and the balance between the generator and discriminator.
2.3 VAE: Variational Autoencoder
VAE (Variational Autoencoder) is a deep learning architecture based on probabilistic generative models.
It maps input data to a probability distribution of latent variables through an encoder, and then generates data from the latent space through a decoder.
Core structure:
- Encoder: the encoder.
- Latent Space: the latent space.
- Decoder: the decoder.
The training objective of VAE usually includes reconstruction loss and a KL divergence regularization term.
Where:
你贴出的是一段关于生成式 AI、NLP 和计算机视觉的教材式内容,末尾停在“八、计算机视觉”标题处。 请告诉我你希望我做什么,例如: 1. 续写“八、计算机视觉”章节 2. 润色、校对或统一术语 3. 压缩成学习笔记/思维导图 4. 转成 Markdown、表格或 Mermaid 图 5. 根据内容出练习题或总结重点 如果没有特别说明,我可以默认帮你续写“八、计算机视觉”部分。Computer Vision (CV) studies how to enable computers to extract information from images and videos, recognize objects, and understand visual content.
Its core task is to convert pixel data into semantically meaningful information.
1. Main Tasks of Computer Vision
| Technical Direction | English Name | Typical Applications |
|---|---|---|
| Image Classification | Image Classification | Determine the category of an image |
| Object Detection | Object Detection | Identify objects and their locations |
| Image Segmentation | Image Segmentation | Pixel-level region segmentation |
| Face Recognition | Face Recognition | Identity verification |
| Pose Estimation | Pose Estimation | Human motion analysis |
| Image Generation | Image Generation | AI image generation |
| Image Enhancement | Image Enhancement | Denoising, super-resolution |
| Video Understanding | Video Understanding | Action recognition, event analysis |
2. Differences Between Typical Vision Tasks
Image classification, object detection, and image segmentation are three concepts that are often confused.
- Image classification: determine which category the entire image belongs to.
- Object detection: identify which objects are in the image and give their locations.
- Image segmentation: further determine which pixel regions correspond to each object.
They differ in the granularity of their understanding of visual information.
For example, in an image containing cars, pedestrians, and a road:
- A classification model may determine that the image belongs to a traffic scene.
- A detection model can mark bounding boxes for cars and pedestrians.
- A segmentation model can further mark the pixel regions of the road, cars, and pedestrians.
3. Development of Vision Models
Computer vision has evolved from manual feature engineering to deep learning.