AI 智能核心词汇

The simulation of human intelligence by machines, enabling them to perform tasks requiring reasoning, learning, and problem-solving — the broadest concept in the field.

A subset of AI where computers learn patterns and make decisions from data without being explicitly programmed, improving performance through accumulated experience.

A branch of machine learning using multi-layered neural networks to automatically learn hierarchical representations directly from raw data such as images, audio, and text.

A computing system inspired by biological neural networks, consisting of interconnected layers of nodes (neurons) that process data through weighted connections and non-linear activation functions.

The iterative process of feeding labeled data through a neural network and adjusting its weights so that the model's output increasingly matches the desired ground-truth results.

The phase where a trained model is used to make predictions or generate outputs on new, unseen data — typically faster and more cost-effective than training, often heavily optimized for deployment.

When a model learns the training data too well — including its noise and quirks — resulting in poor performance on new data and failing to generalize outside the training set.

A model's ability to perform accurately on new, unseen data not part of its training set — the fundamental goal that separates useful models from mere memorization.

A function measuring the discrepancy between a model predictions and the actual ground-truth values, providing the gradient signal used to iteratively update weights during training.

An algorithm computing gradients of the loss with respect to every weight in a neural network by applying the chain rule of calculus layer by layer from output back to input.

An optimization algorithm that iteratively adjusts model parameters in the direction of steepest descent of the loss landscape, seeking a local or global minimum.

An algorithm (e.g., Adam, RMSProp, SGD with momentum) implementing gradient descent, adaptively managing learning rates and momentum to accelerate convergence and escape poor local minima.

Parameters set before training begins (e.g., learning rate, batch size, number of layers) that control the training process and critically affect model performance and convergence speed.

Techniques (L1/L2 penalty, dropout, data augmentation) that discourage overfitting by adding constraints or injecting noise during training, improving the model ability to generalize.

A dense, low-dimensional representation of discrete entities (words, items, nodes) where similar items are positioned close together in the learned vector space, capturing semantic relationships.

The process of breaking raw text into discrete tokens — words, subwords (BPE, WordPiece, SentencePiece), or characters — that a language model can process as numerical input.

A neural network using learnable filters (kernels) that slide across input data to automatically detect spatial hierarchies — the foundational architecture for modern computer vision.

A neural network architecture designed for sequential data, where connections between nodes form directed sequences, allowing information to persist across time steps via hidden states.

A specialized RNN architecture with gating mechanisms (input, forget, output gates) that allow it to learn long-range dependencies and effectively mitigate the vanishing gradient problem.

A neural network architecture introduced in 2017 (Attention Is All You Need) relying entirely on self-attention mechanisms, eliminating recurrence and convolution, enabling massive parallelization across GPUs.

A mechanism computing a weighted sum of all positions in a sequence for each position using learned Query, Key, and Value projections, enabling capture of long-range dependencies regardless of distance.

A neural network with billions to hundreds of billions of parameters, pre-trained on massive amounts of text data, capable of understanding and generating human language at remarkable scale and fluency.

The discipline of crafting and optimizing input prompts to elicit desired outputs from language models — including few-shot prompting, chain-of-thought reasoning, and structured system instructions.

The process of taking a pre-trained model and continuing training (updating weights) on a smaller, task-specific or domain-specific dataset to adapt it for a particular application or behavior.

An architecture that augments a language model generation by retrieving relevant documents from an external knowledge base at inference time, reducing hallucination and improving factual accuracy.

When an LLM generates confident, fluent responses that are factually incorrect, internally inconsistent, or unsupported by its training data — one of the most critical challenges in LLM deployment.

A training methodology that uses human preference labels to train a reward model, then optimizes the policy via reinforcement learning (typically PPO) to align model outputs with human values and intent.

AI systems that can autonomously plan, reason, use tools, and execute multi-step workflows to accomplish complex goals without continuous human intervention — going beyond single prompt-response interactions.

The research field focused on ensuring AI systems behave in ways that are beneficial and consistent with human intentions and values, preventing unintended or harmful outcomes as capabilities increase.

Empirical observations that a model performance improves predictably as the amount of training data, model parameters, and compute increase — often following clean power-law relationships across many orders of magnitude.

A specialized processor originally designed for graphics rendering, now the dominant accelerator for deep learning due to its massively parallel architecture with thousands of cores optimized for matrix/tensor operations.

A Google-designed ASIC optimized for tensor operations that powers large-scale deep learning workloads in Google Cloud and research, offering high throughput for matrix multiplications.

A specialized AI accelerator chip designed to efficiently execute neural network operations (convolution, matrix multiplication) with low power consumption, commonly integrated into mobile SoCs and edge devices.

An integrated circuit that can be reconfigured after manufacturing, offering a flexible hardware platform for custom AI acceleration with lower latency than GPUs for specific workloads.

A measure of computing performance representing one trillion (10^12) floating-point operations per second, commonly used to describe the throughput of GPUs and AI accelerators for training and inference workloads.

Dedicated memory on a GPU used to store model parameters, activations, and intermediate computations during training and inference — a critical bottleneck for large models that determines maximum model size.

The practice of splitting a large model and/or large dataset across multiple GPUs and machines, training in parallel using data parallelism, model parallelism, and pipeline parallelism to overcome single-device limits.

A distributed computing paradigm that processes data near the source (IoT devices, smartphones, local servers) rather than in centralized cloud data centers, reducing latency and enabling real-time AI inference.

An open-source container orchestration platform that automates deployment, scaling, and management of containerized applications across clusters — widely used for serving ML models at scale with automatic load balancing.

The delivery of computing resources (servers, storage, GPU, networking) over the internet on a pay-per-use basis, providing on-demand access to the massive infrastructure required for training and deploying large AI models.

A lightweight, standalone package that encapsulates a software application along with its runtime dependencies and configuration — enabling consistent ML model deployment across development, testing, and production environments.

A high-level, interpreted programming language known for its readable syntax and vast ecosystem of libraries (NumPy, PyTorch, TensorFlow), making it the dominant language for AI research and machine learning development.

Low-level, high-performance programming languages that provide fine-grained control over memory and computation, widely used for implementing AI inference engines, GPU kernels (CUDA), and performance-critical AI serving infrastructure.

A systems programming language emphasizing memory safety, concurrency, and zero-cost abstractions — increasingly adopted for writing safe, high-performance AI inference servers and compiler toolchains (e.g., llama.cpp).

A statically typed, compiled language designed for simplicity and concurrency, popular for building scalable ML serving systems, data pipelines, and cloud-native AI infrastructure due to its excellent concurrency primitives.

A domain-specific language for managing and querying relational databases — essential for feature engineering, data preprocessing, and retrieval in AI/ML pipelines, as well as vector databases that support RAG architectures.

A versatile, interpreted language that powers web browsers and Node.js — increasingly used for deploying lightweight ML inference via TensorFlow.js, ONNX.js, and client-side AI features directly in web applications.

The deployment of AI models directly on edge devices (smartphones, IoT sensors, smart cameras, wearables) enabling real-time inference without cloud connectivity, addressing privacy, latency, and bandwidth constraints.

The development of vehicles capable of sensing and navigating their environment without human input, using a combination of computer vision, sensor fusion (LiDAR, radar), reinforcement learning, and high-definition mapping.

A computing paradigm that leverages quantum mechanical phenomena (superposition, entanglement) to perform certain calculations exponentially faster than classical computers, with potential applications in optimization, drug discovery, and cryptography.