Math to Model: GRPO, PCA Augmentation, and SVD

I’ve been diving deep into some challenging problems on Deep-ML, ranging from classical linear algebra to modern reinforcement learning. Here are my solutions and takeaways for three distinct problems.

1. GRPO Objective Function (Reinforcement Learning)

Implementing the Group Relative Policy Optimization (GRPO) objective from the DeepSeekMath paper was a great refresher on RL math.

Key Insight: The equations look intimidating, but it’s really just a loss function designed to control weight updates. It combines a PPO-style clipped surrogate with a KL divergence penalty to keep the model from “going berserk.”

👉 Code Solution

2. PCA Color Augmentation (Computer Vision)

This is the technique used in the legendary AlexNet. Instead of random noise, we use PCA to find the principal colors in an image and add noise along those semantic axes.

Key Insight: It’s an elegant way to simulate lighting intensity and color temperature changes while preserving the image’s semantic content.

👉 Code Solution

3. SVD of a 2x2 Matrix (Linear Algebra)

Singular Value Decomposition (SVD) is the grandfather of dimensionality reduction.

Key Insight: Implementing this from scratch brought back memories of college linear algebra. It’s beautiful how the eigenvectors of $A^T A$ give us the right singular vectors of $A$.

👉 Code Solution

---

Solving these problems is the best way to keep the mathematical intuition sharp!