Glossary

0-9

1-shot learning 2-stage detector 3D convolution 4D data 5G + AI 6DoF pose estimation 7D representation 8-bit quantization 9-layer network 0-shot learning

A

Artificial Intelligence (AI)Algorithm Attention Autoencoder AGI / Artificial General Intelligence

B

BERT Boosting Backpropagation Batch Normalization Bias

C

Classifier / Classification Cross-Validation Clustering CNN / Convolutional Neural Network Chatbot

D

Discriminative Model Deterministic Model Deep Learning Data Augmentation Deepfake

E

Epoch Encoder Ensemble Learning Explainable AI (XAI)Embedding

F

Foundation Model Forward Propagation Fusion / Multimodal Fusion Feature Extraction Fine-tuning

G

GAN / Generative Adversarial Network Grounding Generative AI Gradient Descent Graph Neural Network (GNN)

H

Hierarchical Model Hyperparameter Hallucination Heuristic Hidden Layer

I

Interpretability Instruction tuning Imbalanced Data Instance / Sample Intelligence Amplification / Augmentation

J

JAX JSONL / JSON-lines Juxtaposition Jittering Joint Embedding

K

K-Shot Learning Kernel Trick KL Divergence (Kullback–Leibler Divergence)Knowledge Distillation K-means Clustering

L

LSTM / Long Short-Term Memory Large Language Model (LLM)Latent Variable Loss Function Learning Rate

M

Multimodal / Multimodality Multi-head Attention Machine Learning (ML)Model Meta-learning

N

NLU / Natural Language Understanding Normalization Neural Network Novelty Detection / Anomaly Detection NLP / Natural Language Processing

O

One-hot Encoding Overfitting Objective Function Optimizer Online Learning

P

Prompt Parameter Pretraining Policy / Reinforcement Learning Policy Pooling

Q

Q-learning Query Queue / Buffer Quantization Quality Estimation

R

RNN / Recurrent Neural Network Representation Learning Retrieval Augmented Generation (RAG)Reinforcement Learning (RL)Regularization

S

Softmax Sampling Supervised Learning Sequence Modeling Self-Supervised Learning

T

Tokenizer Transformer Tuning / Hyperparameter Tuning Transfer Learning Training Data

U

U-Net Uncertainty Estimation Underfitting Universal Approximation Theorem Unsupervised Learning

V

Vision Transformer (ViT)Variational Autoencoder (VAE)Vector Embedding Vanishing / Exploding Gradient Validation Set

W

Weight Decay Word Embedding Workflow Whitening / Whitening Transformation Weak Supervision

X

X-axis / feature axis XLM XLNet XAI / Explainable AI XOR problem

Y

Yoga of AI Y-transform / YUV Yield (model yield / throughput)Y-axis / feature axis YAGNI (You Aren't Gonna Need It)

Z

Z-score Normalization Zero-gradient phenomenon Zero-shot Learning / Zero-shot inference Zero-centric / Zero-bias initialization Zygosity in augmentation

什么是消失/爆炸梯度

消失梯度和爆炸梯度是深度学习和神经网络训练中至关重要的概念。

消失梯度是指在反向传播过程中，梯度变得极其微小，最终导致权重几乎不再更新。这种现象在使用 sigmoid 或 tanh 激活函数的深层网络中尤为常见，从而导致学习速度缓慢或停滞。

与此相反，爆炸梯度是指在反向传播中，梯度异常增大，导致权重更新不稳，模型难以收敛。这通常发生在使用 ReLU 激活函数的多层网络中。

这两种现象显著影响深度学习模型的训练效率和效果。研究者们提出了 LSTM 等多种架构来缓解消失梯度的影响，并采用梯度裁剪等技术来处理爆炸梯度。

随着深度学习的发展，解决这些问题变得越来越重要，新的激活函数和网络设计应运而生，旨在保持梯度的稳定性。