Advanced Topics in Machine Learning (ATML):

Course material and programming assignments for EE-5102 / CS-6304 – Advanced Topics in Machine Learning (Fall 2025, LUMS), covering deep representation learning, interpretability, efficiency, federated systems, domain adaptation, and reinforcement learning for LLMs.

This repository serves as a comprehensive investigation into the frontiers of modern machine learning. Through a series of hypothesis-driven experiments, it explores how deep learning systems generalize, fail, and can be optimized for real-world deployment across diverse environments.

---

🏛️ Core Research Pillars

1. Investigating Inductive Biases

This module explores how architectural designs impose "built-in" assumptions. We contrast the local receptive fields of CNNs with the global self-attention of Vision Transformers (ViTs) and the multimodal alignment of CLIP.

• Shape vs. Texture: Probing why CNNs are "texture-hungry" while ViTs and CLIP align closer to human-shaped perception.
• Generative Trade-offs: Evaluating the latent manifold smoothness of VAEs versus the high-fidelity outputs of GANs.
• Key Insight: ViTs handle spatial disruptions (occlusions/permutations) significantly better than CNNs by leveraging global context.

2. Robustness & Domain Adaptation

Models often fail when test data deviates from the training distribution. This section implements strategies for Domain Adaptation (DA) and Domain Generalization (DG) using the PACS dataset.

• Invariant Learning: Using IRM and Group DRO to minimize reliance on "spurious correlations" (e.g., specific backgrounds).
• Optimization Geometry: Leveraging Sharpness-Aware Minimization (SAM) to find "flatter" minima that generalize better to unseen styles like Sketch or Art.
• Prompt Tuning: Adapting CLIP via CoOp to specialize on new domains while maintaining open-set flexibility.

3. Advanced Model Compression

To deploy models on edge devices, we must reduce their footprint without sacrificing accuracy. This module benchmarks optimization techniques on VGG architectures:

• Pruning: Comparing unstructured (weight-level) vs. structured (channel-level) pruning for hardware-level speedups.
• Quantization: Analyzing the performance "cliff" of INT4 precision and how Quantization-Aware Training (QAT) recovers accuracy.
• Knowledge Distillation: Using CRD and DKD to transfer "dark knowledge" from high-capacity teachers to compact students.

4. Federated Learning & Heterogeneity

This section addresses the challenges of training models on decentralized, "Non-IID" (label-skewed) data without exchanging raw information.

• Client Drift: Measuring how local data skew (simulated via Dirichlet partitioning) causes local models to diverge from the global objective.
• Mitigation Algorithms: Benchmarking FedProx, SCAFFOLD, and FedSAM to stabilize global aggregation.
• Key Finding: Seeking flatter minima (FedSAM) is more effective at handling data heterogeneity than standard regularization penalties.

5. LLM Alignment & Interpretability

The final module explores how we control Large Language Models (LLMs) and investigate their internal representations.

• Alignment Regimes: A head-to-head comparison of DPO, PPO, and GRPO for reducing verbosity bias and reward hacking.
• Decoding Dynamics: Benchmarking how strategies like Top-P (Nucleus) Sampling balance diversity with coherence.
• Mechanistic Interpretability: Using Universal Sparse Autoencoders (USAEs) to identify shared features between ResNet and ViT, testing the Platonic Representation Hypothesis.

---