ML/DL Knowledge Index

מדריך ניווט למערכת הסקילים של ML/DL מקורס Hebrew University.

Quick Reference: Which Skill to Use

| If the task involves... | Use this skill | |------------------------|----------------| | Regression, classification, evaluation metrics | ml-fundamentals | | Random Forest, XGBoost, clustering, PCA, recommender systems | ml-advanced | | TF-IDF, Word2Vec, topic modeling, text similarity | nlp-classical | | Training loops, loss functions, backpropagation | deep-learning-core | | PyTorch tensors, DataLoader, GPU memory | pytorch-mastery | | CNNs, image classification, transfer learning | cnn-vision | | RNN, LSTM, time series, text generation | sequence-models | | Transformers, BERT, HuggingFace, LLMs | transformers-llm | | RAG, embeddings, vector stores, semantic search | rag-retrieval | | LLM APIs, PDF parsing, chunking, function calling | data-pipeline | | LoRA, QLoRA, PEFT, quantization, instruction tuning, RLHF/DPO | fine-tuning-peft | | MLflow, W&B, experiment tracking, hyperparameter tuning, Optuna | mlops-experiment | | SHAP, feature importance, Grad-CAM, error analysis, explainability | model-interpretability | | Reinforcement learning, Q-learning, DQN, PPO, Gymnasium | reinforcement-learning | | GANs, VAE, diffusion models, Stable Diffusion, image generation | generative-models | | Need explanation, "how does X work", teaching | ml-teaching-assistant |

Skill Summaries

ml-fundamentals

Topics: Linear/Logistic Regression, Decision Trees, Ridge/Lasso, train/test split, cross-validation, Precision/Recall/F1, ROC-AUC, feature engineering, encoding, scaling

ml-advanced

Topics: Random Forest, XGBoost, CatBoost, Stacking, K-Means, DBSCAN, Hierarchical clustering, PCA, t-SNE, UMAP, geospatial analysis, Matrix Factorization, NeuMF, recommender systems

nlp-classical

Topics: Bag-of-Words, TF-IDF, Word2Vec, FastText, GloVe, Doc2Vec, LDA topic modeling, Jaccard/Cosine similarity, FuzzyWuzzy, record linkage

deep-learning-core

Topics: Three Pillars (Model, Loss, Optimizer), gradient descent, backpropagation, Adam/SGD, learning rate, Dropout, BatchNorm, MLP architecture, Autoencoders, Denoising AE, latent space

pytorch-mastery

Topics: Tensor creation, broadcasting, NCHW format, Dataset/DataLoader, training loop patterns, CUDA, GPU memory, .to(device), debugging shapes, environment setup, nvidia-smi

cnn-vision

Topics: Convolution, pooling, feature maps, LeNet/ResNet/VGG, transfer learning, fine-tuning, data augmentation, image preprocessing, MNIST, multi-modal networks, image captioning

sequence-models

Topics: RNN formula, hidden state, vanishing gradients, LSTM/GRU, time series forecasting, text generation, language models, sequence classification

transformers-llm

Topics: Self-attention, Transformer architecture, BERT, MLM/NSP, HuggingFace Tokenizer/Trainer/Pipeline, GPT, Claude, Gemini, prompt engineering

rag-retrieval

Topics: Embedding APIs (OpenAI, Gemini, Sentence-Transformers), FAISS, ChromaDB, Pinecone, RAG variants, query rewriting, RAGAS evaluation, hybrid search

data-pipeline

Topics: OpenAI/Gemini/Ollama setup, LiteLLM, pdfplumber, PyMuPDF, OCR, chunking strategies, function calling, LangChain agents, Pydantic validation

fine-tuning-peft

Topics: LoRA, QLoRA, PEFT library, adapter tuning, instruction tuning, quantization (GPTQ, AWQ, GGUF, bitsandbytes), DPO/RLHF alignment, SFTTrainer, TRL, Unsloth, Axolotl, model merging

mlops-experiment

Topics: MLflow, Weights & Biases, TensorBoard, Optuna hyperparameter tuning, model registry, experiment versioning, learning rate schedulers, early stopping, reproducibility

model-interpretability

Topics: SHAP (TreeExplainer, DeepExplainer, KernelExplainer), feature importance (MDI, Permutation), Grad-CAM, LIME, attention visualization, confusion matrix analysis, error analysis pipeline

reinforcement-learning

Topics: MDP, Q-Learning, DQN (experience replay, target network), Policy Gradient (REINFORCE), PPO, Actor-Critic, Stable-Baselines3, Gymnasium environments, reward shaping

generative-models

Topics: GANs (DCGAN, WGAN), VAE (reparameterization trick, KL divergence), Diffusion Models (DDPM), Stable Diffusion, text-to-image, latent space interpolation, conditional generation

ml-teaching-assistant

Topics: Concept explanations with analogies, visual ASCII diagrams, common misconceptions, progressive complexity, "why" questions

Common Cross-Skill Workflows

"I want to build an image classifier"

1. cnn-vision          → Architecture selection, augmentation
2. pytorch-mastery     → Training loop, DataLoader
3. deep-learning-core  → Loss functions, regularization
4. ml-teaching-assistant → If needs explanation

"I want to build a RAG system"

1. rag-retrieval       → Architecture, vector stores
2. data-pipeline       → PDF parsing, chunking
3. transformers-llm    → Embedding models, LLM selection

"I want to do customer segmentation"

1. ml-advanced         → Clustering algorithms (K-Means, DBSCAN)
2. ml-fundamentals     → Feature engineering, evaluation
3. data-pipeline       → Data preprocessing

"I want to classify text"

Option A (Classical): nlp-classical → TF-IDF + sklearn
Option B (Deep): sequence-models → LSTM embeddings
Option C (Modern): transformers-llm → BERT fine-tuning

"I want to build a recommender system"

1. ml-advanced           → Matrix Factorization, NeuMF architecture
2. pytorch-mastery       → Training loop, DataLoader, GPU
3. deep-learning-core    → Loss functions (MSELoss), embedding layers

"I want to fine-tune an LLM"

1. fine-tuning-peft      → LoRA/QLoRA setup, dataset preparation
2. transformers-llm      → HuggingFace Trainer, tokenization
3. mlops-experiment      → Experiment tracking, hyperparameter tuning

"I want to understand why my model predicts X"

1. model-interpretability → SHAP, Grad-CAM, error analysis
2. ml-fundamentals        → Evaluation metrics, confusion matrix

"I want to train an RL agent"

1. reinforcement-learning → Algorithm selection, environment setup
2. pytorch-mastery        → Neural network for policy/value
3. mlops-experiment       → Tracking RL experiments

"I want to generate images"

1. generative-models      → GAN/VAE/Diffusion architecture
2. cnn-vision             → CNN layers, image processing
3. pytorch-mastery        → Training loop, GPU optimization

Custom Models vs LLMs Decision Framework

| Scenario | Use Custom Models | Use LLMs | |----------|------------------|----------| | Narrow tasks (classification, ranking) | Small models can beat LLMs | - | | Domain-specific jargon, frequent updates | Private data, specialized | - | | Large corpus analysis | LLMs can't comprehend many docs | - | | Tabular / Time-series data | LLMs not suited | - | | Recommender systems | Specialized architectures (MF, NeuMF) | - | | Cost / Privacy concerns | LLMs expensive, external APIs | - | | Flexible NL understanding | - | Quick prototyping | | Document generation / summarization | - | Natural strength | | Question answering with RAG | - | With retrieval pipeline | | Function calling / AI agents | - | Tool-augmented LLMs |

Rule of thumb: Start with the simplest model that meets your needs.

Learning Paths

• New to ML? Start with ml-fundamentals → ml-advanced → deep-learning-core
• Deep Learning Track: deep-learning-core → pytorch-mastery → cnn-vision or sequence-models
• NLP Track: nlp-classical → transformers-llm → rag-retrieval
• LLM Engineering: transformers-llm → fine-tuning-peft → mlops-experiment
• Generative AI: deep-learning-core → generative-models → fine-tuning-peft
• RL Track: deep-learning-core → reinforcement-learning
• Production ML: mlops-experiment → model-interpretability → fine-tuning-peft

Reference

• reference/full_topic_index.md - Complete searchable index of all topics