Artificial Intelligence - AI
Computer systems simulating human intelligence, performing tasks typically requiring human intelligence: learning, reasoning, perception, natural language understanding. Core driver of current tech development.
Reference · Glossary
Discover essential AI terminology to master artificial intelligence. This comprehensive dictionary covers 140+ professional terms from machine learning and deep learning to large language models, each with clear definitions.
10 core concepts terms with clear definitions and practical explanations.
Computer systems simulating human intelligence, performing tasks typically requiring human intelligence: learning, reasoning, perception, natural language understanding. Core driver of current tech development.
AI branch enabling computers to learn from data without explicit programming. ML allows AI systems to automatically improve performance, foundation of most modern AI applications.
ML subset using multi-layer neural networks to simulate human brain learning. Achieved breakthroughs in image recognition, natural language processing.
Computational model mimicking human brain neuron structure, consisting of interconnected nodes (neurons). Core architecture of deep learning, handles complex nonlinear relationships.
AI models trained on vast text data, understanding and generating natural language. LLMs like GPT, Claude, Gemini are essential infrastructure for AI applications.
AI technology generating new content (text, images, video, audio). Generative AI analyzes patterns in training data to create entirely new content, greatly expanding creative expression possibilities.
AI branch enabling computers to understand, interpret, and generate human language. NLP includes text analysis, machine translation, sentiment analysis, dialogue systems.
AI branch enabling computers to recognize and understand image and video content. Widely used in image recognition, object detection, face recognition, autonomous driving.
AI systems processing and understanding multiple data types (text, images, audio, video) simultaneously. Enables more comprehensive world understanding and natural interaction.
AI systems autonomously executing tasks, making decisions, interacting with other systems. Understand user intent, plan actions, execute complex tasks. Advanced form of AI applications.
10 machine learning methods terms with clear definitions and practical explanations.
Machine learning approach training models on labeled data. Each input has corresponding target output, model learns input-to-output mapping. Widely used in classification, regression tasks.
Machine learning approach discovering patterns in unlabeled data without predefined categories. Automatically identifies hidden patterns, widely used in clustering, dimensionality reduction.
Machine learning approach where agents learn optimal policies through environment interaction and reward signals. Uses trial-and-error and feedback mechanisms, enabling AI systems to make decisions in dynamic environments. Widely used in games, robotics.
Technique adapting models trained on one task to related tasks. Leverages knowledge from pre-trained models, quickly adapting to new tasks with less data and computation. Important technology for modern AI applications.
Process of pre-training models on large-scale data. Pre-training enables models to learn general language or visual representations, foundation for fine-tuning and task-specific applications. Most modern AI models use pre-training + fine-tuning paradigm.
Further training pre-trained models with domain-specific data. Fine-tuning adapts AI models to specific tasks or domains, improving expertise and accuracy.
Model's ability to perform tasks without task-specific training data. Zero-shot learning leverages general knowledge from pre-trained models, enabling handling of unseen tasks or data during training.
Model's ability to learn new tasks with only few examples. Few-shot learning uses prompt engineering and in-context learning, enabling rapid adaptation to new tasks. Important technology for improving AI system flexibility.
Machine learning approach combining small amounts of labeled data with large amounts of unlabeled data. Semi-supervised learning is particularly useful in scenarios where labeled data is scarce, leveraging unlabeled data to improve model performance.
Machine learning approach generating supervision signals from data itself. Self-supervised learning requires no manual labeling, learning data representations by designing prediction tasks (e.g., predicting masked parts), important method for pre-training.
10 model architecture & technology terms with clear definitions and practical explanations.
Neural network architecture introduced in 2017, using self-attention mechanisms to process sequential data. Foundation for most modern LLMs (GPT, BERT), enabling parallel processing and significantly improving training efficiency.
Core component of Transformer architecture, allowing models to focus on importance of different positions in sequences. Computes relationships between query, key, and value vectors, enabling understanding of long-range dependencies.
Large language model series developed by OpenAI, generating coherent, contextually relevant text through pre-training on vast text data. GPT series (GPT-3, GPT-4, GPT-5) achieved breakthroughs in text generation and dialogue systems.
Pre-trained language model developed by Google, using bidirectional Transformer encoders to understand context. BERT excels in natural language understanding tasks, foundation for many NLP applications.
Generative model learning data distribution by gradually adding and removing noise. Diffusion models achieved breakthroughs in image generation, core technology of Stable Diffusion and similar tools.
Neural network architecture consisting of generator and discriminator, generating high-quality content through adversarial training. Core technology of early AI image generation, widely used in image synthesis and style transfer.
Neural network architecture capable of processing sequential data, maintaining memory of historical information through recurrent connections. RNNs were widely used in natural language processing and time series prediction before Transformers.
RNN variant solving long-distance dependency problems through gating mechanisms. LSTMs better remember long-term information, excel in sequence modeling tasks.
Neural network architecture specialized for processing image data, using convolutional layers to extract local features. CNNs are widely used in image recognition, object detection, and other computer vision tasks.
Neural network architecture using residual connections to solve deep network training difficulties. Residual connections allow direct gradient propagation, making training deeper networks possible.
12 training & optimization terms with clear definitions and practical explanations.
Process of adjusting model parameters using data. Training optimizes loss functions, enabling models to learn patterns and regularities in data. Core step in building AI systems.
Process of using trained models to make predictions on new data. Inference is practical application stage of AI models, requires balancing accuracy and speed to meet real-world application needs.
Learnable weights and biases in AI models. Parameter count (model size) is important metric for model scale. Generally larger parameters mean stronger capabilities, but higher training and inference costs.
Function measuring difference between model predictions and true values. Loss functions guide model training process, optimizing model performance by minimizing loss functions.
Optimization algorithm updating model parameters by computing gradients of loss functions. Gradient descent is core algorithm for neural network training, gradually optimizing models through iterative parameter updates.
Algorithm computing gradients for each parameter in neural networks. Backpropagation propagates error signals from output layer to input layer, making deep learning training possible.
Hyperparameter controlling step size of model parameter updates. Too high learning rate may cause training instability, too low slows training, needs adjustment based on tasks.
Number of samples used in each training iteration. Batch size affects training stability and memory usage, requires balancing efficiency and stability.
Phenomenon where models perform well on training data but poorly on new data. Overfitting indicates models over-memorize training data, lacking generalization ability.
Techniques preventing model overfitting, including L1 regularization, L2 regularization, Dropout. Regularization improves model generalization ability by constraining model complexity.
Technique randomly dropping some neurons during training. Dropout reduces dependencies between neurons, prevents overfitting, improves model generalization ability.
Technique normalizing inputs of each layer. Batch normalization accelerates training process, improves model stability, makes training deeper networks possible.
10 text processing terms with clear definitions and practical explanations.
Basic unit in text processing, can be words, subwords, or characters. Tokens are foundation for AI models processing text, models decompose input text into token sequences. Token count directly affects model input length and processing capability.
Technique mapping discrete objects (words, images) to continuous vector spaces. Embeddings capture semantic relationships between objects, enabling models to understand similarity and associations. Foundation for AI models understanding the world.
Maximum number of tokens a model can process simultaneously. Context window represents model's \"memory\" capacity, determining text length models can understand and process.
Designing and optimizing input prompts to guide AI models toward desired outputs. Key skill for using generative AI tools, directly affects output quality.
Prompting technique encouraging AI models to reason step-by-step before answering. Chain-of-thought prompting improves accuracy for complex tasks by guiding models to show reasoning processes.
AI architecture enhancing answer accuracy by retrieving relevant web content in real-time. Core mechanism of GEO optimization, widely used in AI search engines.
AI technology generating coherent, relevant text content from input prompts. Widely used in content creation, code generation, dialogue systems.
Automatically translating text from one language to another using AI technology. Modern machine translation is based on neural networks, capable of handling complex language structures and context.
AI technology analyzing emotional tendencies expressed in text. Sentiment analysis can identify positive, negative, or neutral emotions in text, widely used in social media monitoring and customer feedback analysis.
Technology identifying and extracting named entities (person names, locations, organization names, etc.) from text. Named entity recognition is foundation for information extraction and knowledge graph construction.
10 image & vision terms with clear definitions and practical explanations.
AI technology generating new images from text descriptions or reference images. Image generation technologies like DALL-E, Midjourney, Stable Diffusion are transforming design and creative industries.
Technology generating images from text descriptions. Text-to-image tools enable non-professionals to create high-quality images, lowering image creation barriers.
Technology generating new images from reference images and text descriptions. Image-to-image allows modifying and creating based on existing images, providing flexible image generation capabilities.
Technology applying artistic style from one image to another. Style transfer can create new images with specific artistic styles, widely used in creative design and artistic creation.
Technology using AI to improve image quality, including resolution enhancement, noise reduction, color optimization. Image enhancement can convert low-quality images to high-definition.
Technology using AI to adjust image lighting effects, including changing light direction, intensity, color, creating different lighting atmospheres. Image relighting is widely used in product photography and design.
Computer vision technology identifying object locations and categories in images. Object detection can locate multiple objects in images, widely used in autonomous driving, security monitoring.
Computer vision technology dividing images into different regions or objects. Image segmentation precisely identifies which region each pixel belongs to, important technology for medical image analysis and autonomous driving.
Computer vision technology identifying and verifying face identities. Face recognition identifies specific individuals by analyzing facial features, widely used in security systems and identity verification.
Computer vision task assigning images to predefined categories. Image classification is fundamental computer vision task, widely used in content moderation, medical diagnosis.
10 video & audio terms with clear definitions and practical explanations.
AI technology generating video content from text descriptions or static images. Video generation automatically creates video scenes, animations, visual effects, greatly simplifying video production.
Technology generating video content from text descriptions. Text-to-video automatically creates video scenes, animations, visual effects, greatly simplifying video production.
Technology generating dynamic videos from static images. Image-to-video converts photos to videos, adds dynamic effects and animations, widely used in content creation and marketing.
Technology generating new videos from reference videos, changing video styles, adding effects, or editing videos. Video-to-video provides powerful video creation capabilities.
Technology converting text to natural, fluent speech. TTS tools generate speech in multiple languages and voices. Widely used in audio content, voice assistants, accessibility services.
Technology converting human speech to text. Speech recognition is important bridge between AI and human interaction. Widely used in voice assistants, subtitle generation, meeting records.
Technology converting speech signals to editable text. STT tools transcribe conversations in real-time, generate subtitles, improving content creation and meeting recording efficiency.
AI technology copying and cloning specific person voices. Learns target voice characteristics, generates highly similar speech. Used in dubbing, personalized voice assistants.
AI tools changing voice timbre, style, and characteristics in real-time. Converts ordinary voices to various interesting effects. Widely used in entertainment, games, privacy protection.
AI technology automatically synchronizing lip movements in videos or images with audio. Lip sync tools automatically adjust speaker mouth movements, expressions, and actions based on audio content. Widely used in multilingual video localization, virtual anchors.
10 evaluation & performance terms with clear definitions and practical explanations.
Proportion of correctly predicted samples to total samples. Accuracy is basic metric for evaluating classification model performance, but may be inaccurate in imbalanced datasets.
Proportion of true positives among samples predicted as positive. Precision measures reliability of model predictions, high precision means more trustworthy predictions.
Proportion of actual positives correctly predicted as positive. Recall measures model's ability to discover positive samples, high recall means fewer missed detections.
Harmonic mean of precision and recall. F1 score comprehensively considers precision and recall, common metric for evaluating classification model performance.
Matrix table showing model classification results. Confusion matrix displays correspondence between predicted and actual results for each category, helps analyze model error patterns.
Model evaluation method dividing dataset into multiple subsets, using different subsets as validation sets in turn. Cross-validation more accurately evaluates model performance, reducing impact of data division.
Experimental method comparing effects of two versions (A and B). A/B testing evaluates which version performs better by randomly assigning users to different versions, important method for optimizing AI systems.
Method testing model performance using standard datasets and evaluation metrics. Benchmarks enable performance comparison between different models, driving AI technology progress.
Process of evaluating AI model performance using various metrics and methods. Model evaluation includes accuracy, speed, resource consumption, ensuring models meet practical application needs.
Techniques improving AI model running speed and efficiency. Performance optimization includes model compression, quantization, pruning, enabling models to run in resource-constrained environments.
10 data & training terms with clear definitions and practical explanations.
Dataset used for training AI models. Quality and quantity of training data directly affect model performance, high-quality training data is foundation for building excellent AI models.
Dataset used for adjusting model hyperparameters and evaluating model performance. Validation data helps select optimal model configurations, prevents overfitting.
Dataset used for final evaluation of model performance. Test data is not used during training, ensuring objectivity and reliability of evaluation results.
Process of adding labels or annotations to training data. Data labeling is key step in supervised learning, labeling quality directly affects model performance.
Technique generating new training samples by transforming existing data. Data augmentation includes rotation, scaling, flipping, increases training data diversity, improves model generalization ability.
Process of identifying and correcting errors, incomplete or inconsistent data in datasets. Data cleaning is important step in data preprocessing, ensuring training data quality.
Process of extracting and constructing useful features from raw data. Feature engineering helps models better understand data, improves model performance.
Process of selecting most relevant features from all features. Feature selection reduces model complexity, improves training efficiency and model performance.
Situation where sample quantities of different categories in dataset vary greatly. Data imbalance may cause models to favor majority classes, requires handling through sampling or weight adjustment.
Technology identifying anomalies or abnormal patterns in data. Anomaly detection can discover data points that don't conform to normal patterns, widely used in fraud detection, fault diagnosis.
10 applications terms with clear definitions and practical explanations.
AI systems capable of natural conversation with users. Dialogue systems understand user intent, generate relevant responses, widely used in customer service, assistants, education.
AI systems recommending content based on user historical behavior and preferences. Recommendation systems predict content users may be interested in by analyzing user data, widely used in e-commerce, video platforms.
Technology optimizing website content and technology to improve search engine rankings. AI technology is used in SEO for content generation, keyword optimization, user experience analysis.
Process of automatically generating text, images, video and other content using AI technology. Content generation improves creation efficiency, reduces content production costs.
Technology automatically generating code from natural language descriptions or templates. Code generation quickly converts programming ideas to executable code, greatly improving development efficiency.
Using AI technology to generate and execute test cases. Automated testing improves software quality and release efficiency, reduces manual testing workload.
Systems providing customer service using AI technology. Intelligent customer service can respond to customer questions 24/7, improves service efficiency and customer satisfaction.
AI technology providing customized suggestions based on user characteristics. Personalized recommendation improves user satisfaction and conversion rates, widely used in e-commerce, content platforms.
Analytical methods using AI technology to predict future trends and outcomes. Predictive analytics predicts future possible events by analyzing historical data, supports decision-making.
Technology using AI to monitor system status in real-time, identifying abnormal situations. Anomaly monitoring can timely discover system problems, improves system reliability.
9 model optimization terms with clear definitions and practical explanations.
Techniques reducing model size and computational requirements. Model compression includes pruning, quantization, knowledge distillation, enabling models to run on resource-constrained devices.
Technology converting model parameters from high precision (e.g., 32-bit floats) to low precision (e.g., 8-bit integers). Model quantization significantly reduces model size and inference time while maintaining good performance.
Technology removing unimportant parameters or connections from models. Model pruning reduces model complexity, improves inference speed, while maintaining model performance as much as possible.
Technology transferring knowledge from large models (teacher models) to small models (student models). Knowledge distillation enables small models to learn capabilities of large models, reducing computational requirements while maintaining performance.
Technology combining predictions from multiple models to improve performance. Model ensemble improves prediction accuracy and stability through averaging or voting.
Training strategy gradually transitioning from simple to complex tasks. Progressive training improves model learning efficiency, enables models to better master complex tasks.
Learning strategy training models in order of increasing difficulty. Curriculum learning simulates human learning process, improves model learning efficiency and final performance.
Technology training models with adversarial examples to improve robustness. Adversarial training enables models to resist malicious attacks, improves model reliability in practical applications.
Training method simultaneously learning multiple related tasks. Multi-task learning improves model performance on multiple tasks through shared representations, improves data utilization efficiency.
9 deployment & inference terms with clear definitions and practical explanations.
Process of deploying trained models to production environments. Model deployment needs to consider performance, stability, scalability, ensuring models can run stably.
Computing mode performing AI inference locally on devices. Edge computing reduces data transmission latency, improves response speed, protects data privacy, suitable for real-time application scenarios.
Computing mode performing AI inference on cloud servers. Cloud inference can leverage powerful computing resources, supports large-scale concurrency, suitable for complex models and batch processing.
Encapsulating AI models as callable service interfaces. Model serving enables models to be called by other systems, supports microservices architecture and distributed deployment.
Techniques reducing AI model inference time. Latency optimization includes model optimization, hardware acceleration, parallel processing, improving response speed for real-time applications.
Techniques improving number of requests AI systems can process per unit time. Throughput optimization improves system processing capacity through batch processing, parallel inference.
Process of managing different versions of AI models. Model versioning supports model rollback, A/B testing, gradual rollout, ensuring safety of model updates.
Process of monitoring deployed model performance and data distribution in real-time. Model monitoring can timely discover model performance degradation, data drift, ensuring models remain effective.
Release strategy gradually promoting new models to more users. Gradual rollout reduces risks, timely discovers and solves problems, ensures stability of model updates.
10 ethics & security terms with clear definitions and practical explanations.
Systematic errors or unfair results produced by AI models. AI bias usually stems from training data bias, may lead to discriminatory decisions, important ethical issue in AI applications.
Principle ensuring AI systems treat different groups fairly. Algorithmic fairness requires models not to produce discriminatory results due to gender, race, age.
AI systems capable of explaining their decision processes and results. Explainable AI helps users understand model reasoning logic, improves transparency and trustworthiness of AI systems.
Carefully designed input samples that can mislead AI models into producing incorrect predictions. Adversarial examples reveal vulnerabilities of AI models, important research area in AI security.
Technology protecting user data from leakage and abuse. Data privacy is important ethical issue in AI applications, requires balancing data utilization and privacy protection.
Technology enabling multiple participants to collaboratively train models while protecting data privacy. Federated learning enables models to learn from distributed data without sharing raw data.
Mathematical framework protecting individual privacy in data analysis. Differential privacy protects attackers from inferring specific individual information from results by adding noise.
Technology protecting AI models from attacks and abuse. Model security includes adversarial attack protection, model watermarking, access control.
Phenomenon where AI models generate seemingly reasonable but actually incorrect information. AI hallucination is common problem in generative AI, needs mitigation through prompt engineering, fact-checking.
Development and usage principles ensuring AI systems are fair, transparent, explainable, secure. Responsible AI requires considering social impact of AI, ensuring AI technology benefits humanity.
9 emerging technologies terms with clear definitions and practical explanations.
AI systems capable of reaching or exceeding human level across various tasks. AGI is long-term goal of AI research, not yet achieved, but related research progress exists.
AI systems capable of interacting with physical environments. Embodied AI combines perception, reasoning, and action capabilities, enabling AI to execute tasks in real world, intersection of robotics and AI.
AI approach combining neural networks and symbolic reasoning. Neuro-symbolic AI combines pattern recognition capabilities of deep learning with logical reasoning capabilities of symbolic systems, improves explainability and reasoning capabilities of AI.
Technology understanding causal relationships between events. Causal inference enables AI systems to understand \"why\" not just \"what\", improves scientificity of decisions.
Technology learning how to learn. Meta-learning enables AI systems to quickly adapt to new tasks, master new skills with few examples, improves AI flexibility.
AI capability learning new knowledge while maintaining existing knowledge. Continual learning enables AI systems to continuously adapt to new environments and tasks, avoids catastrophic forgetting.
Technology learning representations by contrasting similar and dissimilar samples. Contrastive learning enables models to learn meaningful feature representations, widely used in image and text understanding.
Technology training AI models using human feedback. RLHF makes model outputs more aligned with human values through human ratings and preferences, widely used in dialogue system training.
Prompting technique guiding AI models to reason step-by-step. Chain of thought improves model accuracy on complex tasks by showing reasoning processes, makes AI reasoning more transparent.
This glossary covers comprehensive terminology from basic concepts to advanced techniques, helping you better understand professional knowledge in related fields.
Continuously UpdatedThis site uses cookies and similar technologies for analytics, personalized ads (via Google AdSense), and essential functions. By clicking “Accept All”, you consent to our use of cookies. You can reject non-essential cookies by clicking “Reject All”.