Alm

Part B — Kernel Learning (Centered Alignment)

Student name: Pintu singh Student ID: 230105
Assignment: Midsem Part B
Deadline: 12 March 2026, 8 AM

---

📄 Paper Reference

"Algorithms for Learning Kernels Based on Centered Alignment"
Cortes, Mohri, Rostamizadeh — JMLR 2012

---

📁 Repository Structure

partB/
├── task_1_1.ipynb          # Q1: Algorithm steps (ALIGN / ALIGNF)
├── task_1_2.ipynb          # Q1: Key assumptions
├── task_1_3.ipynb          # Q1: Baseline comparison
├── task_2_1.ipynb          # Q2: Toy dataset setup
├── task_2_2.ipynb          # Q2: ALIGNF implementation
├── task_2_3.ipynb          # Q2: Results + plots
├── task_3_1.ipynb          # Q3: Ablation study
├── task_3_2.ipynb          # Q3: Failure mode analysis
├── report.pdf              # Q4: 4-page summary report
├── requirements.txt        # Python dependencies
├── data/
│   └── README.md
├── results/                # Generated plots / figures
│   ├── task2_results.png
│   ├── task3_1_ablation.png
│   └── task3_2_failure.png
└── llm_task_1_1.json       # Q4: LLM interaction logs
    llm_task_1_2.json
    llm_task_1_3.json
    llm_task_2_1.json
    llm_task_2_2.json
    llm_task_2_3.json
    llm_task_3_1.json
    llm_task_3_2.json
    llm_task_4_1.json
    llm_task_4_2.json

---

📝 Task Summary

Task	Description	Marks
Q1 (task_1_1 to 1_3)	Paper understanding — markdown only, no code	25
Q2 (task_2_1 to 2_3)	ALIGN/ALIGNF implementation on toy dataset	40
Q3 (task_3_1 to 3_2)	Ablation study + failure mode analysis	35
Q4	Report PDF + LLM JSON logs	30
Total		130

---

🧠 Core Idea

• Problem: SVM ke liye optimal kernel manually choose karna mushkil hai.
• Solution: Multiple base kernels ko combine karo — unke Centered Alignment (target ke saath similarity) ke basis pe weights automatically seekho.

Two Algorithms:

Algorithm	Approach
ALIGN	Har kernel ka weight independently compute karo (simpler)
ALIGNF	Weights jointly QP solve karke seekho (better performance)

---

⚙️ Setup & Installation

pip install -r requirements.txt

Requirements:

• Python 3.8+
• numpy
• scikit-learn
• matplotlib
• cvxpy (for ALIGNF QP solver)
• jupyter

---

🚀 How to Run

1. Clone the repository:

   git clone https://github.com/Alm
   cd 230105-midsem/partB

2. Install dependencies:

   pip install -r requirements.txt

3. Run notebooks in order:

   task_1_1.ipynb → task_1_2.ipynb → task_1_3.ipynb
   task_2_1.ipynb → task_2_2.ipynb → task_2_3.ipynb
   task_3_1.ipynb → task_3_2.ipynb
   

4. All outputs/plots will be saved in results/

---

📊 Key Results

• ALIGNF outperforms uniform kernel combination on toy dataset
• Ablation study shows centered alignment is the critical component
• Failure mode: algorithm degrades when kernel matrices are highly correlated

---

⚠️ Notes

• Part A (paper: Centered Alignment, JMLR 2012) has been submitted via Google Form — Part B is valid.
• All notebooks must be executed with outputs visible before final GitHub submission.
• Google Form submission link: https://forms.gle/yxfmRprmHDeAzx1C7