Glossary (Lug'at)

ML/AI/MLOps sohasidagi muhim terminlarning inglizcha-o'zbekcha lug'ati. Har termin uchun qisqacha izohva kontekst.

A

• Activation function(faollik funksiyasi) — neural network'da nonlinearity qo'shadigan funksiya (ReLU, Sigmoid, Tanh).
• AdamW — Adam optimizatori + better weight decay; modern default.
• Agent (AI Agent) — LLM + tools + memory; goal'ga erishish uchun ketma-ket harakatlar.
• Anchor box — object detection'da predefined bounding box shape.
• ANN (Approximate Nearest Neighbor) — yaqin vektorlarni tez topish (HNSW, IVF).
• API (Application Programming Interface) — dasturlash interfeysi.
• Async / await — Python'da concurrent operations.
• Attention mechanism — sequence'dagi muhim qismlarga "diqqat" qaratish.
• AUC (Area Under Curve) — ROC curve ostidagi maydon (classification metric).
• AutoGrad — PyTorch'ning avtomatik gradient hisoblash mexanizmi.

B

• Backpropagation — gradient'larni orqaga tarqatish; neural network o'rgatish algoritmi.
• Bagging (Bootstrap Aggregating) — parallel ensemble (Random Forest asosi).
• Batch — bir vaqtda model'ga uzatilgan sample'lar to'plami.
• Batch Normalization (BN) — activation'larni batch ichida normallashtirish.
• Bayesian Optimization — smart hyperparameter qidiruv (Optuna).
• Bias (matematik) — model output'iga qo'shiladigan constant.
• Bias (xulosa) — algoritmda noto'g'ri prediction'larga moyillik.
• Boosting — sequential ensemble (XGBoost, LightGBM).
• BPE (Byte-Pair Encoding) — subword tokenization (GPT, Llama'da).
• Broadcasting — NumPy/PyTorch'da turli shape'dagi tensor'larga operatsiya.

C

• Calibration — model probability'larini ishonchli qilish.
• Canary deployment — yangi versiyani kichik traffic'da sinash.
• Categorical feature — diskret qiymatli feature (city, color).
• Chain-of-Thought (CoT) — LLM'da step-by-step reasoning prompt.
• Checkpoint — model state saqlash (resume training uchun).
• Classification — sample'ni diskret class'larga ajratish.
• Clustering — o'xshashlarni guruhlash (unsupervised).
• CNN (Convolutional NN) — image processing uchun neural network.
• Cold start — yangi user/item haqida data yo'q muammosi.
• Concept drift — input → output relationship vaqt o'tishi bilan o'zgarishi.
• Confusion Matrix — TP, FP, TN, FN ko'rsatadigan jadval.
• Context window — LLM bir vaqtda ko'ra oladigan token soni.
• Cosine similarity — ikki vektor orasidagi cos burchak.
• CRD (Custom Resource Definition) — Kubernetes custom obyekt.
• Cross-encoder — sentence pair'lar uchun classifier (reranking'da).
• Cross-validation (CV) — model'ni bir necha bo'lakda baholash.
• CUDA — NVIDIA GPU'larda parallel computation.

D

• DAG (Directed Acyclic Graph) — Airflow'da workflow ko'rinishi.
• Data augmentation — sun'iy ravishda training data kengaytirish.
• Data drift — input distribution vaqt o'tishi bilan o'zgarishi.
• Data leakage — test/validation data training'ga "sizib o'tishi" (xato).
• DataFrame — Pandas'da tabular data strukturasi.
• DataLoader — PyTorch'da batch yuklash.
• Decision Tree — qoidalar daraxtidan iborat klassik ML algoritmi.
• Deep Learning (DL) — chuqur (ko'p qatlamli) neural network'lar.
• DevOps — software development + operations integratsiyasi.
• Diffusion model — image generation (Stable Diffusion, DALL-E).
• Dimensionality reduction — feature'lar sonini kamaytirish (PCA, t-SNE).
• Docker — application containerization.
• Dropout — overfitting'ni kamaytirish uchun neuron'larni tasodifiy "o'chirish".
• DVC (Data Version Control) — Git for data.

E

• EDA (Exploratory Data Analysis) — ma'lumotlarni tahlil qilish bosqichi.
• Embedding — diskret obyektni dense vektorga aylantirish.
• Encoder-Decoder — translation/summarization arxitekturasi.
• Ensemble — bir nechta model birgalikda.
• Epoch — butun dataset bo'yicha bir martalik training.
• Evaluation — model sifatini o'lchash.
• Evidently AI — drift detection va monitoring tool.

F

• F1 Score — precision va recall'ning harmonic mean.
• FastAPI — modern Python web framework (Pydantic asosida).
• Feature — model input'idagi har bir o'lchov.
• Feature engineering — yangi feature'lar yaratish.
• Feature store — feature'larni saqlash va serve qilish (Feast).
• Few-shot learning — kam misol bilan o'rgatish.
• Fine-tuning — pretrained modelni o'z task'ga moslashtirish.
• Flask — micro web framework (FastAPI'dan oldingi standard).
• F-score — F1 ning umumiy holati (beta parametri bilan).
• Function calling / Tool use — LLM'ga tashqi function'larni chaqirishga ruxsat.

G

• GAN (Generative Adversarial Network) — generator + discriminator.
• Gemini — Google'ning LLM oilasi.
• Generative AI — content yaratuvchi AI (matn, rasm, audio).
• Gini index — Decision Tree'da split quality.
• GitHub Actions — CI/CD platform.
• GPT (Generative Pretrained Transformer) — OpenAI LLM oilasi.
• GPU (Graphics Processing Unit) — parallel computation uchun.
• Gradient — funksiyaning eng tez o'sish yo'nalishi.
• Gradient Boosting — sequential boosting algoritm.
• Gradient Descent — loss'ni minimize qilish algoritmi.
• Grafana — monitoring dashboard.
• GridSearch — hyperparameter exhaustive qidiruv.

H

• Hallucination — LLM'ning ishonchli ko'rinishda noto'g'ri javob berishi.
• Helm — Kubernetes package manager.
• HNSW (Hierarchical Navigable Small Worlds) — fast ANN algorithm.
• HPA (Horizontal Pod Autoscaler) — Kubernetes auto-scaling.
• HuggingFace — ML modellar va datasetlar uchun platform.
• Hybrid search — vector + keyword (BM25) qidiruv.
• HyDE (Hypothetical Document Embeddings) — RAG texnikasi.
• Hyperparameter — training'dan oldin belgilangan parametr (lr, batch).

I

• Image segmentation — pixel-level classification.
• Imbalanced data — class'lar soni teng emas.
• Inference — model bilan prediction qilish.
• Ingress — Kubernetes external HTTP routing.
• Instance segmentation — har object'ga alohida mask.
• Instruction tuning — instructions bilan fine-tuning.
• IoU (Intersection over Union) — object detection metric.

J

• Jupyter Notebook — interactive Python environment.

K

• Keras — high-level NN API (TensorFlow'da).
• K-Fold Cross-validation — dataset'ni K ta foldga bo'lish.
• K-Means — clustering algoritmi.
• KNN (K-Nearest Neighbors) — yaqin K ta sample asosida classification.
• Kubernetes (K8s) — container orchestration.
• Kubeflow — Kubernetes-native ML platform.

L

• L1, L2 regularization — Lasso (L1), Ridge (L2).
• LangChain — LLM application framework.
• LangGraph — stateful multi-agent workflows.
• Langfuse — LLM observability platform.
• LayerNorm — Layer normalization (Transformer'larda).
• Learning rate (lr) — gradient descent qadam kattaligi.
• LightGBM — fast gradient boosting (Microsoft).
• Linear Regression — eng oddiy regression algoritmi.
• LLM (Large Language Model) — katta til modeli.
• LlamaIndex — RAG framework.
• LoRA (Low-Rank Adaptation) — efficient fine-tuning.
• Loss function — model xatosini o'lchaydigan funksiya.

M

• MAE (Mean Absolute Error) — regression metric.
• MAP (mean Average Precision) — object detection metric.
• MAPE (Mean Absolute Percentage Error) — % ko'rinishidagi xato.
• MCP (Model Context Protocol) — Anthropic'ning agent tool standarti.
• MinMaxScaler — feature'larni [0, 1]'ga keltirish.
• MLflow — experiment tracking platform.
• MLOps — ML + DevOps integratsiyasi.
• Model registry — versionlangan modellar saqlash.
• MSE (Mean Squared Error) — regression loss.
• Multi-class classification — 3+ class'lar orasida tanlash.
• Multi-label classification — bir sample'ga bir nechta label.
• Multi-task learning — bir model bir nechta task.

N

• N-gram — N ta consecutive so'zlar.
• Naive Bayes — probabilistic classifier (text uchun mashhur).
• NER (Named Entity Recognition) — matnda nomlangan obyektlar.
• Neural Network (NN) — bir-biriga bog'langan neuronlar tarmog'i.
• NLP (Natural Language Processing) — matn bilan ishlash.
• NMS (Non-Maximum Suppression) — overlapping detection'larni filter.
• Normalization — feature'larni bir xil scale'ga keltirish.
• NumPy — numerical computation library.

O

• One-Hot Encoding — categorical → binary vektor.
• ONNX (Open Neural Network Exchange) — cross-framework model format.
• OpenAI — GPT yaratuvchi kompaniya.
• Optimizer — gradient'ni qanday qo'llash (SGD, Adam, AdamW).
• Optuna — Bayesian hyperparameter tuning.
• Overfitting — model train'da yaxshi, test'da yomon.

P

• Pandas — tabular data manipulation.
• Parameter — modelda o'rganiladigan qiymat (weight).
• PCA (Principal Component Analysis) — dimensionality reduction.
• PEFT (Parameter-Efficient Fine-Tuning) — LoRA, QLoRA va h.k.
• Perceptron — eng oddiy neuron.
• Pipeline — sklearn'da preprocessing + model.
• Pod — Kubernetes'da eng kichik unit.
• Pooling — CNN'da downsampling (MaxPool, AvgPool).
• POS tagging (Part-Of-Speech) — gap bo'laklarini aniqlash.
• Postgres / PostgreSQL — relational database.
• Precision — TP / (TP + FP).
• Prefect — modern workflow orchestrator.
• Pretrained model — katta corpus'da oldindan o'rgatilgan model.
• Prompt — LLM'ga beriladigan input matn.
• Prompt engineering — yaxshi prompt yozish san'ati.
• Prometheus — metrics monitoring system.
• PSI (Population Stability Index) — drift detection metric.
• Pydantic — Python data validation.
• PyTorch — deep learning framework.

Q

• QLoRA — 4-bit quantization + LoRA.
• Qdrant — vector database (Rust).
• Quantization — model precision'ini kamaytirish (8-bit, 4-bit).
• Query — LLM/search'ga beriladigan savol.

R

• R² — coefficient of determination (regression).
• RAG (Retrieval Augmented Generation) — LLM + knowledge retrieval.
• RAGAS — RAG evaluation framework.
• Random Forest — bagging Decision Trees.
• RandomizedSearch — random hyperparameter qidiruv.
• Recall — TP / (TP + FN).
• Recommender system — tavsiya sistemasi.
• ReAct (Reasoning + Acting) — agent pattern.
• Recurrent Neural Network (RNN) — sequence uchun NN.
• Redis — in-memory database.
• Regex (Regular Expression) — pattern matching.
• Regression — uzluksiz qiymat bashorat.
• Regularization — overfitting'ni kamaytirish (L1, L2, Dropout).
• Reranking — search natijalarini qayta tartibga solish.
• REST API — HTTP-based API standard.
• ResNet — skip connection'lari bo'lgan CNN.
• RLHF (Reinforcement Learning from Human Feedback) — LLM alignment.
• RMSE (Root Mean Squared Error) — sqrt(MSE).
• ROC-AUC — Receiver Operating Characteristic Area Under Curve.

S

• SageMaker — AWS ML platform.
• Scaler — feature normalization (Standard, MinMax).
• scikit-learn — Python ML library.
• Self-attention — sequence ichidagi token'lar orasidagi attention.
• Self-supervised learning — labels'siz pretraining.
• Semantic search — meaning-based qidiruv (vector search).
• Sentence Transformer — sentence embeddings.
• SFT (Supervised Fine-Tuning) — instruction'lar bilan fine-tune.
• SGD (Stochastic Gradient Descent) — klassik optimizer.
• SHAP (SHapley Additive exPlanations) — model interpretation.
• Shadow deployment — yangi modelni traffic'siz sinash.
• Sigmoid — activation function (binary class uchun).
• Softmax — multi-class output activation.
• spaCy — NLP library.
• Standardization — (x - mean) / std.
• Streaming — real-time response (SSE, WebSocket).
• Supervised learning — labels bilan o'rganish.
• SVM (Support Vector Machine) — klassik classifier.

T

• Tensor — multi-dimensional array (NumPy ndarray'ning generalizatsiyasi).
• TensorFlow — Google'ning DL framework'i.
• Test set — yakuniy baholash uchun ajratilgan data.
• TF-IDF — text feature representation.
• Threshold — classification decision chegarasi.
• Token — tokenization'dan keyingi atomic unit.
• Tokenizer — matnni token'larga ajratish.
• TorchServe — PyTorch production serving.
• Train set — model o'rganadigan data.
• Transfer learning — pretrained model'ni o'z task'ga qo'llash.
• Transformer — attention-based arxitektura (BERT, GPT).
• Triton — NVIDIA inference server.

U

• Underfitting — model juda oddiy, train'da ham yomon.
• Unicode — character encoding standard.
• Unsupervised learning — labels'siz o'rganish.

V

• Validation set — hyperparameter tuning uchun data.
• Variance — data tarqoqlik darajasi.
• Vector — 1-D array.
• Vector Database — embeddings saqlash va search.
• ViT (Vision Transformer) — rasm uchun Transformer.
• vLLM — fastest LLM inference server.

W

• WandB (Weights & Biases) — experiment tracking.
• Weight — neuron coefficient.
• WebSocket — bidirectional connection.
• Word2Vec — word embedding model.
• Workflow orchestration — task'lar ketma-ketligini boshqarish (Airflow).

X

• XGBoost — popular gradient boosting library.
• XLM-R — multilingual RoBERTa.

Y

• YAML — config fayl formati.
• YOLO (You Only Look Once) — fast object detection.

Z

• Zero-shot learning — pre-existing knowledge bilan misol'siz task.

Asosiy sahifaga qaytish yoki Resurslar ga o'ting.