Glossary

0-9

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

A

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

B

Backpropagation BERT Bias Boosting Batch Normalization

C

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

D

Data Augmentation Deep Learning Deepfake Deterministic Model Discriminative Model

E

Ensemble Learning Encoder Embedding Epoch Explainable AI (XAI)

F

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

G

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

H

Hierarchical Model Heuristic Hallucination Hidden Layer Hyperparameter

I

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

J

Juxtaposition Jittering JAX Joint Embedding JSONL / JSON-lines

K

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

L

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

M

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

N

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

O

Online Learning Objective Function One-hot Encoding Overfitting Optimizer

P

Pretraining Pooling Parameter Policy / Reinforcement Learning Policy Prompt

Q

Q-learning Queue / Buffer Quantization Quality Estimation Query

R

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

S

Self-Supervised Learning Supervised Learning Sequence Modeling Sampling Softmax

T

Tuning / Hyperparameter Tuning Tokenizer Training Data Transfer Learning Transformer

U

Unsupervised Learning Universal Approximation Theorem Underfitting Uncertainty Estimation U-Net

V

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

W

Weight Decay Weak Supervision Whitening / Whitening Transformation Word Embedding Workflow

X

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

Y

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

Z

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

¿Qué es Weight Decay?

Weight Decay es una técnica de regularización ampliamente utilizada en aprendizaje automático y aprendizaje profundo, destinada a prevenir el sobreajuste.

Funciona agregando un término de penalización a la función de pérdida que desalienta los valores de peso grandes, alentando efectivamente al modelo a aprender pesos más pequeños.

Esta técnica es particularmente beneficiosa para modelos complejos y conjuntos de datos de alta dimensión, ya que ayuda al modelo a generalizar mejor cuando encuentra datos no vistos.

Weight Decay a menudo se utiliza junto con otros métodos de regularización, como Dropout, para aumentar la robustez del modelo.

A medida que avanza la tecnología de aprendizaje profundo, Weight Decay puede evolucionar aún más a través de tasas de aprendizaje adaptativas y métodos de optimización más sofisticados.