Day 19-21: Specialized (NLP/CV)

Executive Summary

Domain	Core Task	Standard Feature/Arch
NLP	Text Classification/Translation	Embeddings, Transformers, attention
CV	Image Classification/Detection	Convolutions (CNN), Pooling
Tabular	Forecasting/Fraud	XGBoost, LightGBM

---

文本 1. Natural Language Processing (NLP)

Text Representation

1. Bag-of-Words (BoW): Counting word frequencies. High-dim, sparse.

2. TF-IDF: Weighs words by importance ($TF \times IDF$).

3. Embeddings (Word2Vec/GloVe): Dense vectors where relationship is captured by distance.

Modern NLP: Transformers

• Self-Attention: Allows the model to weigh different words in a sentence differently when processing a token.

• BERT: Encoder-only, good for understanding (Sentiment, NER).

• GPT: Decoder-only, good for generation.

---

2. Computer Vision (CV)

Convolutional Neural Networks (CNNs)

• Filters/Kernels: Learn features like edges, textures, and patterns.

• Pooling: Reduces spatial dimensions (Max Pooling is most common).

• Invariance: CNNs allow for translation invariance (an ear is an ear, wherever it is in the picture).

Vision Transformers (ViT)

Recent breakthrough: Applying the Transformer architecture to patches of images instead of pixels.

---

Interview Questions

1. "What is TF-IDF and why use it over simple counts?"

TF (Term Frequency) measures how often a word appears in a doc. IDF (Inverse Document Frequency) penalizes common words (like "the", "is"). TF-IDF highlights words that are unique/important to a specific document.

2. "Why are CNNs better than MLPs for images?"

MLPs have too many parameters (fully connected) and don't capture spatial locality. CNNs use shared weights (filters) and translation invariance, making them much more efficient and effective for visual patterns.

3. "What is Data Augmentation in the context of CV?"

Artificially increasing the dataset by applying transformations (flips, rotations, crops, color shifts) to the original images. This helps the model generalize better and reduces overfitting.

---

NLP Quick-Check

from sklearn.feature_extraction.text import TfidfVectorizer

vectorizer = TfidfVectorizer(stop_words='english')
X = vectorizer.fit_transform(corpus)