Advanced Deep Learning System

Self-Supervised Learning • Anomaly Detection • Continual Learning

📌 Overview

This project implements an end-to-end Deep Learning pipeline consisting of:

• ✅ Self-Supervised Pretraining using Autoencoders
• ✅ Anomaly Detection via Reconstruction Error
• ✅ Continual Learning to prevent Catastrophic Forgetting
• ✅ Model Evaluation & Performance Comparison

The system demonstrates representation learning, unsupervised anomaly detection, and task adaptation in a modular and scalable setup using PyTorch.

---

🧠 Project Architecture

Pretraining (SSL)
        ↓
Anomaly Detection
        ↓
Continual Learning
        ↓
Evaluation & Comparison

---

🚀 Features

• Self-supervised representation learning (no labels required)
• Reconstruction-error–based anomaly scoring
• Percentile-based anomaly thresholding
• Continual learning with regularization to preserve prior knowledge
• Performance comparison between pretrained and continual models
• Clean modular structure for experimentation

---

🛠️ Tech Stack

• Python 3.x
• PyTorch
• Torchvision
• NumPy
• Matplotlib

---

📂 Project Structure

continual_dl_project/
│
├── pretrain_ssl.py
├── anomaly_detection.py
├── continual_learning.py
├── evaluation_comparison.py
│
├── best_pretrained_model.pth
├── continual_model.pth
│
└── data/

---

🔍 Phase 1: Self-Supervised Pretraining

• Trained an Autoencoder on MNIST dataset.
• Objective: Minimize reconstruction loss.
• Output: best_pretrained_model.pth

Loss Function

Mean Squared Error (MSE)

Output

• Training loss trend (visualized)
• Saved pretrained model checkpoint

---

🚨 Phase 2: Anomaly Detection

• Loaded pretrained model.
• Computed reconstruction error for each sample.
• Applied 95th percentile threshold.
• Samples exceeding threshold flagged as anomalies.

Anomaly Score

Reconstruction Error = Mean((Input - Output)^2)

---

🔁 Phase 3: Continual Learning

• Adapted model to rotated MNIST (new task).
• Implemented EWC-style regularization.
• Prevented catastrophic forgetting.
• Output: continual_model.pth

Continual Objective

Total Loss = New Task Loss + λ * Weight Regularization

---

📊 Phase 4: Evaluation & Comparison

Compared:

• SSL Pretrained Model
• Continual Learning Model

Metric:

• Average Reconstruction Loss (Lower = Better)

Output:

• Bar graph comparison
• Quantitative loss values

---

📈 Results Summary

• SSL successfully learned robust representations.
• Anomaly detection identified high-error samples.
• Continual learning preserved performance close to pretrained baseline.
• Minimal catastrophic forgetting observed.

---

🏆 Key Concepts Demonstrated

• Self-Supervised Learning
• Unsupervised Anomaly Detection
• Elastic Weight Regularization
• Catastrophic Forgetting Mitigation
• Modular Deep Learning Design

---

💡 How to Run

1️⃣ Pretraining

python pretrain_ssl.py

2️⃣ Anomaly Detection

python anomaly_detection.py

3️⃣ Continual Learning

python continual_learning.py

4️⃣ Evaluation

python evaluation_comparison.py

---

📌 Future Improvements

• Add multiple sequential tasks
• Implement full Elastic Weight Consolidation (Fisher-based)
• Add visualization of reconstructed images
• Deploy as an interactive dashboard

---

📜 Conclusion

This project demonstrates a complete deep learning workflow from representation learning to continual adaptation.
It highlights practical implementation of advanced ML concepts used in real-world AI systems.

---

👩‍💻 Author

Isha Singh Rathore Deep Learning Enthusiast | PyTorch Developer