NLP & Computer Vision Applications

NLP Applications

Sentiment Analysis:

# Interview approach
def sentiment_pipeline():
    # 1. Preprocessing
    text = text.lower().strip()
    tokens = tokenize(text)
    
    # 2. Features
    # Option A: TF-IDF + Logistic Regression (baseline)
    # Option B: Word embeddings (Word2Vec, GloVe)
    # Option C: Transformers (BERT, RoBERTa)
    
    # 3. Model
    # Start simple: Logistic Regression on TF-IDF
    # Production: Fine-tuned BERT
    
    return model.predict(features)

Interview Q: "Build sentiment classifier for product reviews" A:

1. Data: Labeled reviews (1-5 stars)
2. Preprocessing: Lowercase, remove punctuation, handle negations ("not good")
3. Baseline: TF-IDF + Logistic Regression
4. Advanced: Fine-tune DistilBERT (faster than BERT)
5. Handle: Sarcasm (hard), emojis, misspellings
6. Evaluation: F1 per class, confusion matrix

Text Classification:

• Spam detection: NB or Logistic Regression (fast, interpretable)
• Intent classification: BERT embeddings
• Named Entity Recognition (NER): BiLSTM-CRF or transformer

Computer Vision Applications

Image Classification:

# Interview approach
def image_classifier():
    # 1. Architecture
    # Transfer learning: Pre-trained ResNet50, EfficientNet
    # Fine-tune top layers on custom data
    
    # 2. Data augmentation
    # Random crop, flip, rotation, color jitter
    
    # 3. Training
    # Adam optimizer, learning rate schedule
    # Early stopping on validation loss
    
    return model

Interview Q: "Classify X-ray images for disease detection" A:

1. Transfer learning: ImageNet pre-trained ResNet50
2. Data augmentation: Rotations, flips (medical images are centered)
3. Class imbalance: Weighted loss (rare diseases)
4. Evaluation: AUC-ROC per disease, sensitivity (recall)
5. Interpretability: GradCAM heatmaps (show radiologist what model sees)
6. Validation: Stratified k-fold, test on different hospitals

Object Detection:

• YOLO: Real-time (autonomous driving, video)
• Faster R-CNN: Higher accuracy (medical, satellite imagery)
• Two-stage: Region proposals → classification

Interview Q: "Detect pedestrians for self-driving car" A:

• Model: YOLOv8 (real-time, 30+ FPS)
• Data: Augment with weather conditions, lighting
• Evaluation: mAP (mean Average Precision), IoU threshold
• Edge cases: Occlusion, night time, small objects
• Deployment: TensorRT optimization for GPU inference

:::