The Architect's Choice

Before you write code, you must choose the architecture. The exam will give you a scenario and ask: "Which algorithm usually performs best?"

---

1. The Algorithm Cheat Sheet

Data Type	Problem	Best Architecture	Why?
Tabular	Classification/Regression	XGBoost / LightGBM (Trees)	Handles unscaled features well; interpretable.
Tabular	Very Complex / Massive	Deep Neural Network (Wide & Deep)	Can learn crossed features.
Image	Object Detection	CNN (ResNet, EfficientNet)	Spatial invariance (a cat is a cat even if rotated).
Text	Translation / Summary	Transformer (BERT, T5, Gemini)	Attention mechanism handles long-range context.
Time Series	Forecasting	LSTM / ARIMA+	Remembers history (sequence dependence).
Recommendation	User Preference	Two-Tower Model / Matrix Factorization	Handles sparse data (users usually only buy 1% of items).

---

2. Transfer Learning

Exam Rule: Never train from scratch if you can Fine-Tune.

• Training from Scratch: Requires millions of images. Takes weeks.
• Transfer Learning: Take a Google model (trained on ImageNet), freeze the early layers, and just retrain the last layer on your images. Requires ~100 images. Takes minutes.

Vertex AI Model Garden is the library where you find these base models.

---

3. Designing for Interpretability

Sometimes accuracy isn't the goal. In Banking and Health, Explainability is the law.

• White Box Models: Linear Regression, Decision Trees. (Easy to explain: "You were denied because Age < 18").
• Black Box Models: Deep Neural Networks. (Hard to explain).

If the exam scenario says "Must be fully interpretable by non-technical auditors," avoid Deep Learning. Use BigQuery ML Linear Regression or Boosted Trees.

---

4. Loss Functions

You must tell the model what "Success" looks like.

Problem	Loss Function	Metric
Binary Classification	Binary Cross-Entropy (Log Loss)	Accuracy, AUC, Precision/Recall
Multi-Class	Categorical Cross-Entropy	F1-Score
Regression	MSE (Mean Squared Error)	RMSE, MAE
Outlier-Heavy Regression	Huber Loss / MAE	MAE is robust to outliers; MSE punished them too hard.

---

5. Visualizing the "Wide & Deep" Model

This is a specific Google architecture often tested. It combines memorization (Wide) with generalization (Deep).

graph TD
    Input[Input Features]
    
    subgraph "Wide (Memorization)"
    Linear[Linear Model]
    end
    
    subgraph "Deep (Generalization)"
    Dense1[Dense Layer] --> Dense2[Dense Layer]
    end
    
    Input --> Linear
    Input --> Dense1
    
    Linear --> Add[Combine]
    Dense2 --> Add
    Add --> Output[Sigmoid Output]

---

6. Summary

• Tabular: Trees (XGBoost).
• Perceptual (Vision/Text): Deep Learning (CNN/Transformers).
• Scarcity: Use Transfer Learning.
• Regulation: Use Interpretable Models.

In the next lesson, we scale up. How do we train these models on Terabytes of data? Distributed Training.

---

Knowledge Check