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

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

B

BERT Boosting Bias Backpropagation Batch Normalization

C

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

D

Deepfake Discriminative Model Data Augmentation Deep Learning Deterministic Model

E

Epoch Explainable AI (XAI)Embedding Encoder Ensemble Learning

F

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

G

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

H

Hierarchical Model Hallucination Heuristic Hyperparameter Hidden Layer

I

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

J

JAX JSONL / JSON-lines Juxtaposition Jittering Joint Embedding

K

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

L

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

M

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

N

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

O

One-hot Encoding Optimizer Objective Function Online Learning Overfitting

P

Prompt Pooling Parameter Policy / Reinforcement Learning Policy Pretraining

Q

Q-learning Quality Estimation Query Quantization Queue / Buffer

R

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

S

Softmax Sampling Sequence Modeling Supervised Learning Self-Supervised Learning

T

Tokenizer Transformer Training Data Tuning / Hyperparameter Tuning Transfer Learning

U

U-Net Uncertainty Estimation Underfitting Unsupervised Learning Universal Approximation Theorem

V

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

W

Weight Decay Workflow Weak Supervision Whitening / Whitening Transformation Word Embedding

X

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

Y

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

Z

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

什麼是消失/爆炸梯度

消失梯度和爆炸梯度是深度學習和神經網絡訓練中至關重要的概念。

消失梯度指在反向傳播過程中，梯度變得極其微小，最終導致權重幾乎不再更新。這種現象在使用 sigmoid 或 tanh 激活函數的深層網絡中尤為常見，從而導致學習速度緩慢或停滯。

與此相反，爆炸梯度則是指在反向傳播中，梯度異常增大，導致權重更新不穩，模型難以收斂。這通常發生在使用 ReLU 激活函數的多層網絡中。

這兩種現象顯著影響深度學習模型的訓練效率和效果。研究者們提出了 LSTM 等多種架構來緩解消失梯度的影響，並採用梯度裁剪等技術來處理爆炸梯度。

隨著深度學習的發展，解決這些問題變得越來越重要，新的激活函數和網絡設計應運而生，旨在保持梯度的穩定性。