AI导论：概念框架与学习路径

究竟什么是人工智能？机器真的能够拥有智能吗？——从图灵1950年对"机器能思考吗？"这一问题的操作性定义，到今天能够生成图像和文本、诊断疾病、击败围棋冠军的AI，理解从"机器能思考吗？"这一哲学追问到现代人工智能产业的完整链条，是进入AI领域的第一步。本教程将沿着这条主线逐章展开：从智能的定义出发，建立AI-ML-DL的分层框架，深入神经网络原理与技术演进，分析机器学习三种范式的本质区别，最终落脚于真实的行业应用与可操作的学习路径。

1. 人工智能的核心问题：图灵测试与智能的操作性定义

"机器能思考吗？"

1950年，艾伦·图灵在他开创性的论文《计算机械与智能》（Computing Machinery and Intelligence）中提出了一个看似简单却影响深远的question：机器能思考吗？ Computing Machinery and Intelligence 图灵并没有直接回答这个哲学问题；相反，他提出了一个巧妙的替代方案——模拟游戏（Imitation Game）。这一思想为随后的人工智能研究奠定了方法论基础，可以说是该领域的第一个严肃的学术提案。

模仿游戏的三人版本

原始设置涉及三位参与者：男性A、女性B，以及审讯者C（任意性别）。审讯者与A和B身处不同房间，通过电传打字机进行纯文本交流——所有对话均为纯文本，没有语音、语调或视觉线索。审讯者必须根据对话判断"X是A，Y是B"还是"X是B，Y是A"。B的策略是诚实回答以帮助审讯者；A的策略是故意误导。

在图灵提出的测试版本中：由机器取代A，我们检查审讯者做出错误判断的频率是否与和人类对弈时相当。如果机器能够骗过审讯者，使其无法区分，则认为该机器具有智能。

下图展示了这个测试的基本流程：

flowchart LR
    I["Interrogator C"] -->|Send Question| H["Human B"]
    I -->|Send Question| M["Machine"]
    H -->|Return Text Answer| I
    M -->|Return Text Answer| I
    I -->|"Judgment: Which is Human?"| Result["Conclusion"]

The core logic of this design is behavioralist operationalization: instead of asking whether the machine "is truly thinking," we focus on its external behavior — being indistinguishable equals having intelligence. This is a pragmatic philosophical concession and a precursor to modern AI evaluation.

Nine Categories of Objections and Turing's Responses

In Section 6 of his paper, Turing systematically discusses nine main objections. Here are the most representative:

The Consciousness Argument (Most Famous Objection): That a machine can fool the interrogator does not mean it is truly "thinking" — it may have intelligence but no inner experience. Searle's (John Searle) 1980 Chinese Room argument is a classic representative of this viewpoint: a computer that only knows symbol manipulation rules but does not understand Chinese can still fool the interrogator through symbol manipulation, but this does not mean it understands Chinese.

Lady Lovelace's Objection: Machines can only do what humans know how to instruct them to do; they cannot truly create.

The Mathematical Objection: Gödel's theorem shows that discrete-state machines have undecidable problems; human reason is equally limited by Gödel's theorem, and humans might also fail to answer questions the machine fails on.

Turing himself admitted: "I have no very convincing positive arguments to support my views." This shows that his argument is essentially a speculative hypothesis, not a proof.

Historical Significance and Limitations of the Turing Test

The historical significance of the Turing test is that it transforms an unanswerable philosophical question into an operable experimental design, providing a methodological starting point for subsequent AI research. However, it also has obvious limitations:

• Behaviorist Trap: Passing the test ≠ truly understanding or having consciousness. Symbol manipulation is not equivalent to understanding — this is the core insight of the Chinese Room argument.
• Text Conversation Only: It misses important dimensions of intelligence such as perception, motor ability, and emotion.
• Human-Centric Bias: Using "imitating humans" as the standard may exclude non-human forms of intelligence.

Modern AI research rarely aims to pass the Turing test; instead, it focuses on specific capability evaluations (Benchmarks) — comparing human performance on specific tasks. This may be the most profound impact of the Turing test: it inaugurated a new research paradigm.

2. The AI Landscape: The Hierarchical Relationship of AI-ML-DL and the Global Competition Landscape

Three-Layer Nested Relationship

The first step in understanding the full scope of the AI field is to clarify the relationship between three core concepts:

Artificial Intelligence (AI)
    └── Machine Learning (ML)
            └── Deep Learning (DL)

Artificial Intelligence (AI) is the broadest concept, referring to systems created to perform tasks that typically require human intelligence, including four core capabilities: perception, reasoning, learning, and decision-making.

Machine Learning (ML) is a subset of AI that enables computers to learn from data without explicit programming — the core idea is to use various methods to make predictions based on data.

Deep Learning (DL) is a subset of ML that uses neural networks (especially deep networks) for representation learning, automatically learning hierarchical features from raw data.

A common misconception is equating AI with deep learning. In reality, deep learning is just one of many paths in AI; many AI systems (such as rule-based expert systems) do not use neural networks.

Global AI Competition Landscape: Let the Data Speak

Understanding the scale and competitive landscape of the AI field requires authoritative data. The 2025 AI Index Report published by Stanford HAI provides a global AI development panorama as of the end of 2024.

Caption: 2024 global major countries' notable AI model releases comparison. The United States leads significantly with 40 models, China has 15, and Europe (major countries combined) has 3. Data source: Stanford HAI AI Index Report 2025.

Caption: The US-China model performance gap is rapidly narrowing. Taking MMLU as an example, the gap was 17.5 percentage points in 2023, narrowing to 0.3 percentage points in 2024; HumanEval narrowed from 31.6 percentage points to 3.7 percentage points. This indicates that Chinese model quality is rapidly catching up. Data source: Stanford HAI AI Index Report 2025.

Key Data Points:

• Global AI private investment reached $252.3 billion in 2024 (26% year-over-year increase), with generative AI private investment at $33.9 billion (18.7% year-over-year increase, 8.5x that of 2022)
• 78% of surveyed enterprises reported using AI in 2024 (55% in 2023)
• However, only about one-third (~38%¹) report having launched AI at scale

This data reveals an important trend: AI adoption is growing rapidly, but truly achieving large-scale deployment remains a minority. This suggests there is a significant gap between AI "being usable" and "being used well."

Major Subfields

The AI field contains several core subfields:

• Computer Vision (CV): Image classification, object detection, semantic segmentation, etc.
• Natural Language Processing (NLP): Machine translation, text generation, question-answering systems, etc.
• Reinforcement Learning (RL): Game AI, robot control, decision optimization
• Speech Recognition and Synthesis: Speech-to-text, voice assistants, etc.

Understanding the relationships between these subfields helps find your direction of interest in AI learning.

3. Neural Network Basics: From Single Neurons to the Universal Approximation Theorem

From Functions to Neural Networks

Understanding neural networks starts from the most basic concept of functions. The function $y = f(x)$ defines a mapping rule from input to output: $x$ is the independent variable (input), and $y$ is the dependent variable (output).

The core task of neural networks is learning this mapping function. The simplest neural network unit — a neuron — performs the following computation:

$$z = x_1 w_1 + x_2 w_2 + \cdots + x_K w_K + b$$

Where $w$ is the weight (weights) and $b$ is the bias. Weights determine the degree of influence each input has on the output, while bias provides additional regulatory freedom.

Fully Connected Feedforward Neural Network (FFNN)

Organizing multiple neurons into layers forms a Fully Connected Feedforward Neural Network (FFNN). Information flows from the input layer to the output layer, with each neuron in each layer connected to every neuron in the next layer.

Caption: Fully connected feedforward neural network: fully connected structure from input layer to output layer

When the network contains multiple hidden layers, it is called a Deep Neural Network (DNN). As the number of hidden layers increases, the network's learning capacity also strengthens — each layer extracts more abstract features based on the representations from the previous layer.

Universal Approximation Theorem: Why Neural Networks Work

Understanding why neural networks work requires an important theoretical result — the Universal Approximation Theorem.

This theorem was first established by Hornik et al. in 1989, and its core meaning is: given a nonlinear activation function, a feedforward neural network with a single hidden layer (containing a sufficiently large number of neurons) can approximate any continuous function on a compact set to any desired accuracy.

In layperson's terms: even a relatively simple single-layer network, as long as it has enough neurons, can theoretically represent any complex input-output mapping relationship. This explains why neural networks have powerful modeling capabilities — they provide the possibility of approximating arbitrarily complex functions.

LeNet-5: An Early Success Story of Deep Representation Learning

One of the best examples for understanding the power of neural networks is LeNet-5, proposed by LeCun et al. in 1998. This is a 7-layer convolutional neural network that achieved an error rate below 1% on handwritten digit recognition (MNIST). LeNet-5

Key design elements of LeNet-5 include local connectivity and weight sharing: each convolution kernel only focuses on a local region of the input, greatly reducing the number of parameters; the same convolution kernel shares weights at different positions, enabling the network to detect the same pattern appearing anywhere. This design was inspired by the working principles of biological visual systems — neurons in the visual cortex also have local receptive fields.

LeNet-5's success demonstrated the enormous potential of deep representation learning: stacking multiple processing layers allows the network to automatically learn hierarchical feature representations from low-level to high-level, which is difficult for traditional machine learning methods to achieve.

4. Deep Learning Evolution Milestones: CNN→Transformer→LLM Technology Leap

Seven Key Algorithms in the Past Decade

In the past decade (2013-2024), several key technological breakthroughs have emerged in the deep learning field. Arranged chronologically, the seven most influential algorithms are:

ResNet: Making Extremely Deep Networks Trainable

A core problem faced by deep networks is gradient vanishing and gradient exploding: as the number of layers increases, gradients exponentially decay or inflate during backpropagation, making training difficult.

Caption: The core of ResNet residual network is the dual-path structure of residual blocks. Input passes through two parallel paths — the residual mapping path (containing two convolutional layers) and the shortcut identity mapping path — merged at the addition node. Shortcuts allow gradients to propagate directly, effectively alleviating the gradient vanishing problem in deep networks.

The core idea of residual learning is: instead of directly learning the underlying mapping $H(x)$, learn the residual $F(x) = H(x) - x$. When the underlying mapping is already optimal, learning a residual of zero is much easier than learning an identity mapping from scratch.

ResNet (He et al., 2015) is one of the most cited works in deep learning history (247,440 citations), proving that extremely deep networks can be effectively trained through residual connections. ResNet

Transformer: A Paradigm Shift in Parallel Processing

In 2017, Vaswani et al. published "Attention Is All You Need," proposing the Transformer architecture, which was epoch-making in the deep learning field (144,501 citations).

Caption: Transformer encoder-decoder overall architecture. The left encoder processes the input sequence, and the right decoder generates the output. The core is the self-attention layer — each position can simultaneously attend to all other positions in the sequence, achieving true parallel sequence processing and eliminating sequential dependencies in recurrent neural networks.

Traditional RNN/LSTM requires sequential processing of sequences and cannot parallelize; the Transformer uses the Self-Attention mechanism to process the entire sequence in parallel, with each token able to simultaneously attend to all other positions in the sequence. The Query-Key-Value vector interaction is the core of the attention mechanism: Query is the query vector of the current position, Key is the key vector of all positions, and Value is the value vector of all positions. The dot product matching degree between Query and Key determines the attention weights.

This brings two key advantages: training speed is greatly improved; simultaneously, it better captures long-range dependencies.

From CNN to Transformer to LLM

There is a clear脉络 in technological evolution:

• CNN Era (around 2012): Convolutional neural networks broke through in image tasks, with AlexNet winning the ImageNet competition by a significant margin
• Transformer Era (2017 to present): Self-attention mechanism extended from NLP to vision and other fields; Vision Transformer (ViT) proved that Transformers can directly process image patch sequences
• LLM Era (2020 to present): Transformer-based large language models (GPT, BERT, LLaMA, Claude, etc.) have emerged, demonstrating the powerful capabilities of the large-scale pre-training + fine-tuning paradigm

Understanding this evolutionary path helps grasp where the current capabilities of AI technology come from and why certain methods are more effective than others.

5. The Three Paradigms of Machine Learning: Essential Differences Between Supervised, Unsupervised, and Reinforcement Learning, and Selection Criteria

Supervised Learning: Learning with Labels

Supervised Learning is the most common machine learning paradigm, with the core being learning a function $f: X \rightarrow Y$ from labeled data pairs $D = {(x_1, y_1), (x_2, y_2), ..., (x_n, y_n)}$.

Typical scenarios:

• Email spam classification: input is email text, output is spam/normal labels
• Disease diagnosis: input is medical images or test data, output is benign/malignant labels
• Price prediction: input is housing features, output is price values

Supervised learning requires a large number of labeled (x, y) data pairs, typically ranging from hundreds to millions. Typical algorithms include SVM, decision trees, neural networks (for classification/regression), Logistic Regression, GBT (gradient boosting trees), etc.

Unsupervised Learning: Discovering Data Structure

Unsupervised Learning uses only unlabeled data ${x_1, x_2, ..., x_n}$, with the goal of discovering the internal distribution structure $P(x)$, rather than predicting specific labels.

Typical scenarios:

• Customer segmentation: clustering customers into different groups
• Anomaly detection: identifying data points that do not conform to normal patterns
• Data dimensionality reduction: compressing high-dimensional data into low-dimensional space

Typical algorithms include K-Means, DBSCAN (clustering); PCA, t-SNE, UMAP (dimensionality reduction); Autoencoder (density estimation/generative modeling).

Reinforcement Learning: Learning Strategies Through Rewards

The core of Reinforcement Learning is that an agent performs actions in an environment, receives rewards, and learns a strategy $\pi(a|s)$ to maximize cumulative rewards.

Typical scenarios:

• Game AI: AlphaGo learns Go strategy through self-play
• Robot control: agents learn to execute action sequences in the physical world
• RLHF alignment: fine-tuning large language models with human feedback signals

Reinforcement learning does not require a fixed dataset; instead, it generates (state, action, reward) sequences through interaction with the environment. The core challenge is the exploration-exploitation trade-off: agents need to balance trying new actions (exploration) and exploiting known optimal actions (exploitation).

Paradigm Comparison and Decision-Making

Below is a decision flowchart to help select the appropriate paradigm in real problems:

flowchart TD
    Start["Start"] --> Q1{"Are there labeled (x,y) data pairs?"}
    Q1 -->|Yes| Q2{"Is labeled data sufficient?"}
    Q2 -->|Yes| SL["Supervised Learning"]
    Q2 -->|No| Q3{"Is there unlabeled data?"}
    Q3 -->|Yes| SSL["Semi-supervised/Self-supervised Pretraining + Fine-tuning"]
    Q3 -->|No| More["Get More Data"]
    Q1 -->|No| Q4{"Is the goal to discover data structure?"}
    Q4 -->|Yes| UL["Unsupervised Learning"]
    Q4 -->|No| Q5{"Does it involve sequential decision-making and rewards?"}
    Q5 -->|Yes| RL["Reinforcement Learning"]
    Q5 -->|No| Rethink["Rethink Problem Definition"]

Supervised/unsupervised/reinforcement learning are not mutually exclusive — they are often combined in real-world applications. For example, Go AI first uses supervised learning to imitate human game records, then uses reinforcement learning to improve itself; large language models first use self-supervised learning for pretraining, then use RLHF to align with human preferences.

6. Core Mechanisms of Neural Networks: Activation Functions, Loss Functions, Regularization, and Gradient Descent

Activation Functions: Introducing Nonlinearity

The computation of each layer in a neural network is essentially matrix multiplication: $z = Wx + b$. Without activation functions, no matter how many layers the network has, it is only a superposition of multiple linear transformations, ultimately equivalent to a single linear transformation — which means the network cannot approximate nonlinear functions.

The role of activation functions is to introduce nonlinearity into the network, enabling it to learn complex patterns.

Common activation functions:

ReLU (Rectified Linear Unit) is the default choice for most deep learning problems because its computation is extremely simple (only need to compare with 0), and the gradient in the positive region is always 1, effectively alleviating the gradient vanishing problem. ReLU

Leaky ReLU is a variant of ReLU that allows small gradients for negative inputs, solving the "dying ReLU" problem — when a neuron's output is always 0, the gradient is also always 0, preventing updates.

Loss Functions: Measuring the Gap Between Predictions and Ground Truth

Loss functions measure the difference between predicted and true values, guiding the optimization process. The smaller the loss, the more accurate the model's predictions.

Mean Squared Error (MSE): $$MSE = \frac{1}{n} \sum_{i=1}^{n} (y_i - \hat{y}_i)^2$$

MSE penalizes larger errors more heavily (due to the square term) and is sensitive to outliers, mainly used for regression tasks.

Cross-Entropy: $$H(y, \hat{y}) = -\sum_i y_i \log \hat{y}_i$$

Cross-entropy is suitable for classification tasks, penalizing confident but incorrect predictions more heavily, encouraging the model to output probability distributions close to the true label distribution. In deep learning, cross-entropy is usually combined with the Softmax layer, forming the classic classification loss combination of Softmax + Cross-Entropy.

Regularization: Preventing Overfitting

When a model performs well on training data but poorly on unseen data, it is called overfitting — the model has memorized the noise in the training data rather than learning the underlying patterns.

Regularization prevents overfitting by adding an extra penalty term to the loss function. Regularization penalizes parameters, keeping their values small.

Source Note: The above L1/L2/Elastic Net regularization comparison table comes from a user-uploaded course PPT (parsed/content.md); ref-005 deep learning survey does not have a dedicated chapter discussing this content.

The addition of regularization terms changes the optimization objective: originally minimizing prediction error, now minimizing prediction error + penalty for large weights. This guides the model to learn in a simpler direction, improving generalization ability.

Gradient Descent: The Core Algorithm for Parameter Optimization

Neural networks iteratively optimize parameters through Gradient Descent (GD), with the core idea being to gradually update parameters in the negative direction of the gradient until the gradient approaches zero (finding a local or global optimum).

Basic Steps:

1. Initialize the values of weights $w$
2. Compute the gradient $\frac{\partial L}{\partial w}$
3. Update weights: $w \leftarrow w - \eta \frac{\partial L}{\partial w}$
4. Repeat the above steps until the gradient is sufficiently small

Where $\eta$ is the learning rate, controlling the step size of each update.

flowchart TD
    A["Initialize weights w"] --> B["Compute loss L"]
    B --> C["Compute gradient ∂L/∂w"]
    C --> D["Update weights w ← w - η·∂L/∂w"]
    D --> E{"Is gradient small enough?"}
    E -->|No| B
    E -->|Yes| F["End"]

Backpropagation is the core algorithm for computing gradients, calculating gradients through the chain rule (intuitive explanation of chain rule see supporting PPT²). When training DNNs using gradient-based learning and backpropagation, the algorithm traverses from the final output layer to the initial layer, computing gradients and updating weights layer by layer.

In actual training, there are also various optimization algorithms to improve the original gradient descent:

• SGD (Stochastic Gradient Descent): Computes gradients using one sample or a small batch of samples each time, accelerating convergence
• Adam: Combines the advantages of AdaGrad and RMSProp, maintaining first-order and second-order moment adaptive estimates, and is one of the most popular optimizers currently

7. Current State of AI Industry Applications and Learning Paths: A Comprehensive Guide from Research to Deployment

AI Deployment Status: The Gap Between Promise and Reality

Understanding the true value of AI technology requires facing the current state of enterprise AI adoption. McKinsey's 2025 survey reveals an important phenomenon: 88% of enterprises have routinely used AI in at least one business function, but only about 38% have truly achieved large-scale deployment.

Another key data point comes from the AI Agent field:

• 62% of enterprises are trying AI Agents (autonomous systems capable of planning and executing multi-step workflows)
• However, only 23% have scaled these in production environments; 39% are still in the experimental stage
• Only about 6% of enterprises are classified as "high-performing organizations" — AI contributes more than 5% to their EBIT

This indicates a significant gap between AI technology's potential and its actual deployment effectiveness. Many promises, limited delivery — but this is a rapidly evolving field.

Industry Distribution and Real Challenges

AI penetration in industries is uneven:

• Technology/Telecom (34%), Financial Services (28%), Healthcare and Pharmaceuticals (26%) are leaders in AI Agent adoption
• Healthcare is the fastest-growing industry (23 percentage points year-over-year increase), driven by drug discovery and clinical decision support
• Marketing and Sales has the highest adoption of generative AI (42%), followed by product development (38%) and service operations (35%)

Real challenges in deployment are equally severe:

• 51% of enterprises report having experienced negative AI consequences, with inaccuracy/hallucination problems accounting for 44%, data privacy 31%, and intellectual property issues 27%
• Enterprise size is strongly correlated with AI scaling success: about half of enterprises with annual revenues over $5 billion have achieved AI scaling, compared to only 22% for smaller enterprises

Learning Paths: From Beginner to Advanced

Learning path suggestions based on the preceding content:

Phase 1: Foundation Building (1-3 months)

• Linear algebra basics: matrix operations, vector spaces, eigenvalue decomposition
• Probability and statistics basics: probability distributions, Bayes' theorem, hypothesis testing
• Python programming: familiar with NumPy, Pandas, Matplotlib

Phase 2: Machine Learning Basics (3-6 months)

• Core supervised learning algorithms: linear regression, logistic regression, SVM, decision trees
• Unsupervised learning: K-Means, PCA
• Evaluation methods: cross-validation, bias-variance tradeoff, handling overfitting

Phase 3: Deep Learning (6-9 months)

• Neural network principles: backpropagation, activation functions, loss functions
• Core architectures: CNN (images), RNN/LSTM (sequences), Transformer (mainstream)
• Common frameworks: PyTorch or TensorFlow

Phase 4: Vertical Direction Deepening (9+ months)

• Computer vision direction: object detection, semantic segmentation, image generation
• Natural language processing direction: text classification, sequence-to-sequence, prompt engineering
• Reinforcement learning direction: basic theory, PPO, DQN

Industry Insights: Which Directions Are Worth Investing In

Looking at several notable trends from current market demand:

• Foundation models concentrated in few large companies: GPT/Claude/Gemini three major model families occupy the vast majority of market share; learners do not need to train models from scratch but need to understand how to build application capabilities based on these models
• Security and transparency are industry weaknesses: among 30 highly agentic products, only 4 disclose any security assessments; there is great demand for security talent
• Enterprise-grade Agents are an emerging hotspot: workflow automation, enterprise system integration, and other directions are growing rapidly

References

The core references cited in this topic are as follows; it is recommended to read the originals for deeper learning:

1. Computing Machinery and Intelligence — Alan Turing, Mind, Vol. 59, No. 236 (1950), pp. 433-460
2. Gradient-Based Learning Applied to Document Recognition — LeCun et al., Proceedings of the IEEE (1998)
3. A Decade of Deep Learning: A Survey on The Magnificent Seven — arXiv:2412.16188
4. The AI Index Report 2025 — Stanford HAI, 2025
5. A Comprehensive Review of Deep Learning: Architectures, Recent Advances, and Applications — Information 2024, MDPI
6. Supervised, Unsupervised, and Reinforcement Learning — EngineersOfAI Documentation, 2024
7. The State of AI in 2025: Agents, Innovation, and Transformation — McKinsey & Company, November 2025
8. AI Agent Index 2025 — arXiv:2602.17753, ACM FAccT 2026