Artificial intelligence may appear to be an intelligent system capable of understanding language, recognizing images, analyzing problems, and generating content, but from a technical implementation perspective, it is not made up of any single algorithm.
A complete AI system usually requires multiple key elements to work together. Among them, the three most fundamental and core elements are:
- Data: Provides learning material for the model and determines what information the model can access and learn from.
- Algorithm: Defines how the model learns, how it represents information, and how it performs prediction and generation tasks.
- Compute: Provides the computational resources required for model training and inference, and determines the scale of models and data the system can handle.
Together, these three elements form the foundation of AI development.
Whether it is traditional machine learning, deep learning, or today's large language models and multimodal AI, all depend on the collaboration of data, algorithms, and compute.
However, it should be noted that the three are not simply independent of one another. Data quality affects the learning effectiveness of the algorithm, algorithm structure affects compute requirements, and compute in turn affects model scale, training methods, and the technical routes that can be realized.
This article systematically analyzes the core components, working mechanisms, and interrelationships of artificial intelligence from the three dimensions of data, algorithms, and compute, as well as their specific manifestations in modern AI systems.
1. Why Is AI Inseparable from Data, Algorithms, and Compute?
Traditional software development usually follows explicit rules:
Developers write programs; the programs execute operations according to inputs and predefined logic, and ultimately produce output.
For example:
def calculate_discount(price, discount):
return price * (1 - discount)
result = calculate_discount(100, 0.2)
print(result)
The calculation rules here are explicitly written by the developer.
However, for complex tasks such as image recognition, speech recognition, and natural language understanding, it is difficult to cover all cases by manually writing rules.
For example, determining whether there is a cat in an image:
- The cat may be at different angles.
- The lighting and background may differ.
- The cat's color, body shape, and posture may differ.
- The image may contain occlusion, blur, and noise.
If relying entirely on manual rules, one would need to write a large number of complex conditions, and it would still be difficult to adapt to unseen situations.
Machine learning takes a different approach:
Instead of having developers write all judgment rules entirely by hand, data and learning algorithms are provided so that the model can automatically learn patterns from the data.
A simplified machine learning process can be represented as:
flowchart TB
A["Prepare training data"] --> B["Select learning algorithm"]
B --> C["Perform model training"]
C --> D["Use compute to perform calculations"]
D --> E["Obtain the trained model"]
E --> F["Input new data"]
F --> G["Model prediction"]
G --> H["Output 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:#F5F8F4,stroke:#B5CBBB,color:#40594A
style E fill:#DCEFE5,stroke:#4F8B70,color:#24563F
style F fill:#F5F8F4,stroke:#B5CBBB,color:#40594A
style G fill:#EAF3EC,stroke:#83A995,color:#315D49
style H fill:#DCEFE5,stroke:#4F8B70,color:#24563F
In this process:
- Data determines which samples and patterns the model can learn.
- Algorithms determine how the model learns from the data.
- Compute is responsible for performing the calculations required for training and inference.
The three can be understood as:
| Core element | Core question | Main role |
|---|---|---|
| Data | What to learn? | Provides learning material and information |
| Algorithm | How to learn? | Defines the learning method and computation process |
| Compute | What resources are used for computation? | Supports model training and operation |
Together, the three determine the capability boundaries of an AI system, but the final performance is also affected by factors such as model architecture, training objectives, engineering implementation, evaluation methods, and deployment environment.
2. Data: The Foundation of AI Learning
Data is one of the most important foundations of artificial intelligence.
Machine learning models need to learn features, patterns, and regularities from data. Without suitable data, algorithms cannot effectively learn the target task.
But more data is not always better.
The quality, distribution, accuracy, representativeness, and labeling method of data all affect the performance of the final model.
2.1 What Types of Data Are There in AI?
Depending on the organization form and content of the data, AI data can be divided into multiple types.
1. Structured data
It has a clear data structure and is usually organized in forms such as tables and database records.
For example:
| User ID | Age | Spending Amount | Churned |
|---|---|---|---|
| 1001 | 25 | 1200 | No |
| 1002 | 38 | 3500 | Yes |
| 1003 | 31 | 2100 | No |
Common applications:
- User churn prediction.
- Credit risk assessment.
- Sales forecasting.
- Recommender systems.
- Financial risk control.
2. Unstructured Data
It has no fixed tabular structure and usually requires models to extract features.
For example:
- Text.
- Images.
- Audio.
- Video.
- Documents.
Common applications:
- Natural language processing.
- Image recognition.
- Speech recognition.
- Video analysis.
- Multimodal understanding.
3. Semi-structured Data
It has a certain organizational structure but does not necessarily follow a strict relational table structure.
For example:
- JSON.
- XML.
- Log files.
- HTML.
- Some document data.
In real AI systems, these data are also often used as training data, retrieval data, or model input.
2.2 Classifying Data by Supervision Method
In machine learning, data can also be classified according to whether it contains labels.
| Data Type | Characteristics | Typical Applications |
|---|---|---|
| Labeled data | Contains inputs and target labels | Supervised learning |
| Unlabeled data | Only inputs, with no human-annotated target labels | Unsupervised learning |
| Weakly labeled data | Labels are incomplete or insufficiently precise | Weakly supervised learning |
| Preference data | Contains human preferences over different outcomes | Model alignment |
| Interaction data | Contains feedback generated by users or the environment | Reinforcement learning, recommender systems |
For example:
Labeled data:
Image: a picture of a cat
Label: cat
Unlabeled data:
Image: a picture of an animal
Label: none
Preference data:
Question: How should I learn Python?
Answer A: It is recommended to start with variables and data types.
Answer B: Dive straight into complex asynchronous programming.
Human preference: Answer A
Different types of data correspond to different training objectives and learning methods.
2.3 Training Data, Validation Data, and Test Data
In machine learning projects, data usually needs to be split into different datasets.
| Dataset | Main Purpose | Participates in Parameter Training |
|---|---|---|
| Training Set | Learn model parameters | Yes |
| Validation Set | Tune hyperparameters, select models | Usually not directly involved in gradient training |
| Test Set | Evaluate the model's generalization ability | No |
Typical data splitting workflow:
flowchart TB
A["Raw dataset"] --> B["Data cleaning and preprocessing"]
B --> C["Dataset splitting"]
C --> D["Training set"]
C --> E["Validation set"]
C --> F["Test set"]
D --> G["Model parameter training"]
E --> H["Model selection and hyperparameter tuning"]
G --> H
H --> I["Finalize model"]
F --> J["Independent test evaluation"]
I --> J
style A fill:#DCEFE5,stroke:#4F8B70,color:#24563F
style B fill:#EAF3EC,stroke:#83A995,color:#315D49
style C fill:#F7F4E8,stroke:#C7B77A,color:#675A2D
style D fill:#EAF3EC,stroke:#83A995,color:#315D49
style E fill:#F5F8F4,stroke:#B5CBBB,color:#40594A
style F fill:#F5F8F4,stroke:#B5CBBB,color:#40594A
style G fill:#EAF3EC,stroke:#83A995,color:#315D49
style H fill:#EAF3EC,stroke:#83A995,color:#315D49
style I fill:#DCEFE5,stroke:#4F8B70,color:#24563F
style J fill:#DCEFE5,stroke:#4F8B70,color:#24563F
Here is an easily overlooked issue: data splitting must avoid data leakage.
For example:
- Samples from the test set are used for training in advance.
- Features containing test set information are used.
- The entire dataset is preprocessed before data splitting.
- Highly correlated records from the same user or the same event appear in both the training set and the test set.
These issues can make evaluation results overly optimistic and fail to truly reflect the model's performance on unseen data.
For time-series forecasting, special attention must also be paid to temporal order; data must not be randomly shuffled before splitting.
2.4 Why is data quality more important than data quantity?
Data quantity affects the range of samples the model can learn from, but data quality determines whether the information the model is exposed to is reliable.
Common data quality problems include:
1. Data errors
For example:
- Ages are negative.
- Monetary units are inconsistent.
- Images and labels are mismatched.
- Text contains garbled characters.
2. Data duplication
The same sample appears repeatedly, which may cause the model to over-focus on certain data patterns.
3. Missing data
Some fields or features lack valid values, which may affect training results.
4. Data bias
Training data cannot represent real-world usage scenarios.
For example, an image recognition model is mainly trained on images taken during the day but needs to run in nighttime environments.
5. Label noise
Manual annotation may contain errors, ambiguity, and inconsistency.
For example, a blurry image is labeled as different categories by different annotators.
6. Changes in data distribution
The data distribution during model training changes relative to the data distribution in the actual deployment environment.
This type of issue is usually called Distribution Shift (distribution shift).
2.5 Relationship between data quality and model performance
A model's final performance is not simply determined by data volume.
Data quality can be understood as a combination of multiple factors:
Where:
- A: Accuracy.
- C: Completeness.
- R: Representativeness.
- D: Data diversity.
- L: Label quality.
This is only a conceptual expression, not a universal formula for calculating data quality.
For different tasks, the importance of each factor also varies.
For example:
- Medical imaging tasks place particular emphasis on annotation accuracy.
- Autonomous driving tasks focus on scenario coverage and long-tail data.
- Large language models focus on corpus quality, coverage, and duplicate contamination.
- Recommender systems focus on the timeliness of behavioral data and user distribution.
2.6 What kind of data do large language models need?
Large language models usually require large-scale, multi-domain, multi-source text and other modality data.
Common data sources include:
- Public web pages.
- Books and documents.
- Academic papers.
- Technical documentation.
- Code repositories.
- Dialogue data.
- Manually written instruction data.
- Human preference feedback data.
The purpose of data at different training stages also differs.
| Training stage | Main data | Main purpose |
|---|---|---|
| Pre-training | Large-scale text, code, and multimodal data | Learn basic language and knowledge patterns |
| Supervised fine-tuning (SFT) | Instructions paired with high-quality responses | Learn to follow instructions and task formats |
| Preference alignment | Human preference comparison data | Optimize response quality and behavior |
| Domain fine-tuning | Domain-specific data | Strengthen capabilities in specialized fields |
Note that modern large-model training does not necessarily follow the fixed order above; it may also combine synthetic data, continued pre-training, reinforcement learning, and other training methods.
3. Algorithms: How AI Learns and Processes Information
If data answers the question of "what to learn," then algorithms answer the question of "how to learn."
Algorithms determine the mathematical methods, network structures, and optimization approaches a model uses to learn patterns from data.
Different algorithms suit different problems.
For example:
- Classification problems: determine which category a sample belongs to.
- Regression problems: predict a continuous numerical value.
- Clustering problems: discover natural groupings in data.
- Generation problems: generate new content based on input or existing information.
- Reinforcement learning: learn decision-making policies through environment interaction and reward signals.
3.1 Main Types of Machine Learning Algorithms
Machine learning algorithms can usually be classified by how they learn.
1. Supervised Learning
Learning the mapping between input and target output from labeled data.
Common algorithms:
- Linear regression.
- Logistic regression.
- Decision trees.
- Random forests.
- Support vector machines (SVM).
- XGBoost.
- Neural networks.
Typical applications:
- Spam classification.
- House price prediction.
- Customer churn prediction.
- Risk identification.
2. Unsupervised Learning
Discovering structures, distributions, or latent patterns in unlabeled data.
Common algorithms:
- K-Means.
- DBSCAN.
- Hierarchical clustering.
- PCA.
- Autoencoders.
Typical applications:
- User segmentation.
- Anomaly detection.
- Dimensionality reduction.
- Feature representation learning.
3. Semi-supervised Learning
Training on a combination of a small amount of labeled data and a large amount of unlabeled data.
Suitable for scenarios where manual labeling is costly.
4. Self-supervised Learning
Constructing training objectives from the data itself, without requiring manually provided labels for every sample.
For example, large language models learn language patterns by predicting the next Token.
Self-supervised learning is an important technical approach in the pre-training of modern large models.
5. Reinforcement Learning
An agent learns a policy through interaction with an environment, guided by reward signals.
Typical applications:
- Game AI.
- Robot control.
- Automated decision-making.
- Certain post-training and alignment tasks for large models.
3.2 Algorithms and Models Are Not the Same Concept
These two concepts are often used interchangeably in technical discussions, but they are not exactly the same.
| Concept | Meaning | Example |
|---|---|---|
| Algorithm | Defines the method of learning and computation | Gradient descent, backpropagation, K-Means |
| Model | The parameterized function or structure obtained after learning | A trained neural network |
| Model architecture | Defines the structure of the model | Transformer, CNN, RNN |
| Parameters | Learnable values obtained through model training | Weights, biases |
| Hyperparameters | Configurations set before or during training | Learning rate, batch size, number of layers |
For example:
Transformer is a model architecture.
Adam is an optimization algorithm.
Only after training do the specific parameters and structure combine to form a model instance that can be used for inference.
3.3 How Do Algorithms Enable Models to Learn from Data?
Taking supervised learning as an example, the model needs to learn a mapping function from input to output:
Where:
- x: the input data.
- \hat{y}: the model's prediction result.
- f_\theta: the model function determined by the parameters \theta.
The goal of model training is to find suitable parameters:
Where:
- y: true target.
- L: loss function.
- \theta^*: optimal or approximately optimal parameters obtained through training.
Actual training usually computes the average loss over the entire training dataset:
This means that the algorithm needs to continuously adjust parameters through an optimization process so that the model's predictions are as close as possible to the training targets.
3.4 Why can neural networks learn complex patterns?
Traditional machine learning algorithms usually rely on specific model assumptions or feature design.
Deep learning, by contrast, uses multilayer neural networks to learn complex feature representations through nonlinear transformations between layers.
A simple neural network layer can be expressed as:
Where:
- x: input.
- W: weight matrix.
- b: bias.
- f: activation function.
- h: output features.
Multilayer networks can be further composed:
Each layer can learn feature representations at different levels.
For example, in image recognition:
- Shallow layers may extract edges and textures.
- Middle layers may combine local shapes.
- Deep layers may form more complex visual representations.
Actual neural network structures are far more complex than this simple example.
3.5 Core components of the algorithm
A typical deep learning training algorithm usually involves the following parts:
| Component | Main responsibility |
|---|---|
| Model architecture | Define the computation graph and how parameters are organized |
| Forward propagation | Compute model output based on input |
| Loss function | Measure the difference between predictions and training targets |
| Backpropagation | Compute gradients of the loss with respect to parameters |
| Optimizer | Update parameters based on gradients |
| Regularization | Reduce the risk of overfitting or constrain model behavior |
| Learning rate scheduling | Control the update magnitude at different training stages |
Overall training process:
flowchart TB
A["Training samples"] --> B["Model forward pass"]
B --> C["Obtain prediction result"]
C --> D["Compute loss"]
D --> E["Backpropagation"]
E --> F["Compute parameter gradients"]
F --> G["Optimizer updates parameters"]
G --> H{"Is training finished?"}
H -->|"No"| A
H -->|"Yes"| I["Obtain trained model"]
style A fill:#DCEFE5,stroke:#4F8B70,color:#24563F
style B fill:#EAF3EC,stroke:#83A995,color:#315D49
style C fill:#F5F8F4,stroke:#B5CBBB,color:#40594A
style D fill:#F7F4E8,stroke:#C7B77A,color:#675A2D
style E fill:#EAF3EC,stroke:#83A995,color:#315D49
style F fill:#F5F8F4,stroke:#B5CBBB,color:#40594A
style G fill:#DCEFE5,stroke:#4F8B70,color:#24563F
style H fill:#F7F4E8,stroke:#C7B77A,color:#675A2D
style I fill:#DCEFE5,stroke:#4F8B70,color:#24563F
3.6 Algorithms determine the capability boundaries of models
Different algorithms and model architectures are suitable for handling different types of problems.
| Technical approach | Main characteristics | Typical applications |
|---|---|---|
| Linear models | Simple structure, relatively strong interpretability | Regression, classification |
| Decision trees and ensemble learning | Can handle complex structured features | Risk control, forecasting |
| CNN | Good at extracting local spatial features | Image recognition |
| RNN, LSTM, GRU | Model sequential information | Time-series analysis, early NLP |
| Transformer | Good at modeling sequential relationships and context | LLMs, vision models |
| GAN | Adversarial training through a generator and a discriminator | Image generation |
| Diffusion | Generation through progressive denoising | Image and video generation |
| Reinforcement learning algorithms | Learn policies through rewards and interaction | Decision-making, control |
This needs to be made clear:
More complex algorithms and architectures are not necessarily better.
If the data scale is limited and the task is simple, traditional machine learning models may be easier to train, deploy, and maintain than large deep learning models.
Technology selection needs to comprehensively consider accuracy, cost, latency, interpretability, and actual business requirements.
4. Computing Power: The Computing Resources That Support AI Operations
If data is the learning material and algorithms are the learning methods, then computing power is the computational resources required to complete these computing tasks.
AI model training and inference essentially involve a large number of mathematical operations.
Deep learning models, in particular, need to frequently perform:
- Matrix multiplication.
- Vector operations.
- Convolution computations.
- Attention computations.
- Gradient computations.
- Parameter updates.
As model and data scales continue to increase, the demand for computing resources also grows rapidly.
4.1 What Hardware Mainly Provides AI Computing Power?
AI computing usually relies on CPUs, GPUs, dedicated AI accelerators, and related storage and interconnect devices.
1. CPU (Central Processing Unit)
The central processing unit is good at general-purpose computing and complex control logic.
In AI systems, CPUs are usually responsible for:
- Data preprocessing.
- Task scheduling.
- Program control.
- Data loading.
- Coordination between the CPU and accelerators.
2. GPU (Graphics Processing Unit)
The graphics processing unit has a large number of parallel computing units and is suitable for large-scale parallel numerical computation.
Matrix operations and tensor computations in deep learning have high parallelism, so GPUs have become important hardware for training and inference.
3. NPU (Neural Processing Unit)
A neural network processor specially optimized for neural network computation.
NPUs from different vendors may differ considerably in instruction sets, architecture, operator support, and software ecosystems.
4. TPU (Tensor Processing Unit)
A series of tensor processors introduced by Google and specifically optimized for machine learning computation.
5. Other AI accelerators
This category also includes dedicated chips designed for data centers, edge devices, and specific inference tasks.
4.2 What Are the Differences Between CPUs, GPUs, and AI Accelerators?
| Hardware | Main characteristics | Typical uses in AI |
|---|---|---|
| CPU | Strong generality, good at complex control | Data processing, scheduling, some inference |
| GPU | Large-scale parallel computing | Model training, model inference |
| NPU | Optimized for neural networks | On-device AI, edge inference |
| TPU | Optimized for tensor computation | Large-scale machine learning computing |
| FPGA | Programmable hardware logic | Specific computation acceleration, low-latency processing |
| ASIC | Customized for specific tasks | Dedicated AI inference and computing |
It should be noted:
GPUs do not mean that all computations are faster than CPUs.
If the task scale is very small, the control logic is complex, or data movement costs are high, the parallel advantages of GPUs may not be fully utilized.
Hardware performance needs to be evaluated in the context of actual workloads.
4.3 Why Is AI Particularly Dependent on GPUs?
Many computations in deep learning can be converted into matrix operations.
For example:
Suppose:
Then the output is:
The computation involves a large number of multiplication and addition operations.
This type of computation has high parallelism, allowing different elements or computation blocks to be assigned to multiple computing units for simultaneous processing.
The large number of parallel computing units in GPUs makes them very suitable for this type of workload.
In addition, modern GPUs provide dedicated computing units optimized for matrix multiplication, such as Tensor Cores in NVIDIA GPUs.
These hardware capabilities significantly improve the efficiency of deep learning model training and inference.
4.4 Computing Power Is Not Just Computing Speed
When evaluating AI computing power, one cannot look only at the GPU model or theoretical FLOPS.
Multiple factors also need to be considered:
| Metric | Meaning | Impact |
|---|---|---|
| FLOPS | Floating-point operations per second | Theoretical computing capability |
| VRAM capacity | High-speed memory capacity available to the GPU | Whether it can hold the model and intermediate data |
| VRAM bandwidth | Data transfer capability between the GPU and VRAM | Data read and write speed |
| System memory capacity | Size of system memory | Data loading and preprocessing capability |
| Interconnect bandwidth | Data transfer capability between multiple devices | Distributed training efficiency |
| Storage performance | Data read speed | Data loading efficiency |
| Power consumption | Electricity required for computation | Operating costs and cooling requirements |