Deep Learning Systems Lab

This repository serves as a centralized workspace for deep learning research, experimental prototypes, and system-level explorations. It includes implementations across multiple domains such as adversarial robustness, generative modeling, computer vision, and optimization-driven learning systems.

The focus is not just on model performance, but also on understanding internal behavior, robustness, and architectural trade-offs.

---

Repository Structure

Adversarial Robustness

• CNN_training_fgsm_proof.ipynb
• FGSM_proof_CNN_training.ipynb
  Experiments demonstrating Fast Gradient Sign Method (FGSM) attacks and their impact on CNN training dynamics, along with robustness evaluation.

---

Generative Models

• DCGAN_celebfaces.ipynb
  Implementation of Deep Convolutional GAN trained on CelebFaces dataset for high-quality face generation.

---

EEG & Attention Mechanisms

• EEG_band_attention_block.py
  Custom attention block tailored for EEG signal processing, focusing on band-specific feature learning.

---

Neuroevolution / Optimization

• Neural_Genetic_Algorithm.ipynb
  Exploration of genetic algorithms for optimizing neural network parameters and architectures.

---

Deepfake Detection

• RESNET_DENSENET_deepfake_detection.ipynb
• VIT_Deepfake_detection.ipynb
• Stress_test_deepfake_model.py
  Comparative study of CNN-based (ResNet, DenseNet) and Transformer-based (ViT) architectures for deepfake detection, including robustness and stress testing.

---

Systems & Scheduling

• Task_Scheduling_Hierarchical.py
  Hierarchical scheduling strategies, potentially applicable to distributed ML workloads or resource-aware training pipelines.

---

Key Themes

• Adversarial machine learning and robustness
• Vision models (CNNs, Transformers)
• Generative modeling (GANs)
• Bio-signal processing (EEG)
• Optimization via evolutionary strategies
• Systems-level experimentation and performance considerations

---

Design Philosophy

• Experiment-first approach: rapid prototyping and validation
• System awareness: focus on computational and architectural efficiency
• Modularity: reusable components where applicable
• Exploration-driven: includes both research-grade and exploratory work

---

Requirements

Typical dependencies across notebooks and scripts:

• Python 3.8+
• PyTorch / TensorFlow (varies per notebook)
• NumPy, Pandas
• Matplotlib / Seaborn
• OpenCV (for vision tasks)

Install per project as needed.

---

Usage

Each notebook/script is self-contained.
Recommended workflow:

1. Open the relevant notebook
2. Install dependencies if missing
3. Execute sequentially
4. Modify hyperparameters or architecture for experimentation

---

Notes

• This repository is not structured as a production package.
• Code quality and structure may vary across experiments.
• Some scripts are exploratory and may require cleanup or refactoring.

---

Future Directions

• Standardizing experiment pipelines
• Adding benchmarking and evaluation suites
• Integrating experiment tracking (e.g., WandB)
• Expanding adversarial defense techniques
• Scaling models and datasets

---