Adaptive Entropy-Guided Universal Domain Adaptation (AEG-UDA)

A comprehensive course project on Universal Domain Adaptation (UDA) that proposes a new framework — AEG-UDA — to enhance the adaptability and robustness of models in non-overlapping and partially overlapping label spaces. This method integrates dynamic thresholding, entropy-guided pseudo-labeling, and a novel dynamic rejection loss to outperform existing approaches like DANCE across multiple adaptation scenarios.

🧠 Course: Advanced Topics in Machine Learning (ATML - CS)
📅 Semester: Fall 2024 🎓 Institution: LUMS
👨‍💻 Contributors: Muhammad Saad Haroon, Jawad Saeed, Daanish Uddin Khan

🧠 Key Contributions

1. Dynamic Adaptive Threshold
   Automatically adjusts entropy thresholds during training to better separate confident and ambiguous target samples.

   

2. Entropy-Guided Pseudo-Labeling (EGPL)
   Assigns soft pseudo-labels with entropy-aware weighting to confident target predictions.

   

3. Dynamic Rejection Loss (DRL)
   Penalizes uncertain (high-entropy) predictions dynamically to reduce noise and confusion.

4. Supports All UDA Variants
   

---

🗃️ Dataset Preparation

To begin, download the Office-31 dataset, which is required for all experiments. The dataset can be obtained from the following link:
📥 Office-31 Dataset

Prepare the dataset in the following directory structure:

data/
├── amazon/Images/
├── dslr/Images/
├── webcam/Images/

---

Once downloaded:

1. Place the zipped dataset in your Google Drive
2. Ensure it is accessible during training by mounting your Google Drive in Google Colab

---

🚀 How to Run

1. Open the corresponding file in Google Colab. This can be:

   • DANCE emulated script.
   • AEG-UDA script.

2. Mount your Google Drive to access the dataset:

   from google.colab import drive
   drive.mount('/content/drive')

3. Connect to a GPU and run the cells sequentially

---

🖥️ Shell Commands for Training

# ODA
!sh script/run_office_obda.sh 0 /content/DANCE/configs/office-train-config_ODA.yaml

# OPDA
!sh script/run_office_opda.sh 0 /content/DANCE/configs/office-train-config_OPDA.yaml

# CDA
!sh script/run_office_cls.sh 0 /content/DANCE/configs/office-train-config_CDA.yaml

# PDA
!sh script/run_office_cls.sh 0 /content/DANCE/configs/office-train-config_PDA.yaml

---

📊 Results & Evaluation

All experimental results, performance comparisons, and visualizations are detailed in the final report:

📄 Project_Report_Final.pdf

Key insights:

• AEG-UDA achieves competitive performance across CDA, ODA, PDA, and OPDA settings
• Improved generalization via entropy-aware pseudo-labeling
• Faster inference than DANCE in several scenarios
• Stronger cluster compactness in t-SNE visualizations

---

📈 Evaluation Metrics

• Test Accuracy
• Mean Per-Class Accuracy (MPCA)
• Inference Time
• Loss Curves
• t-SNE Feature Visualization

---

🔭 Future Work

• Add mixed-precision training for faster convergence
• Extend DRL with uncertainty-based dynamic weighting
• Adapt for streaming/real-time domain adaptation
• Scale to larger, multi-modal datasets

---

📚 Citation

If you use or reference this work, please cite:

@project{AEG-UDA2025,
  title = {Adaptive Entropy-Guided Universal Domain Adaptation (AEG-UDA)},
  note = {Developed as part of the CS-6304: Advanced Topics in Machine Learning course at LUMS, 2025.}
}

---

“In a shifting world, the best models don’t memorize — they adapt.”