Week 16 – Deep Learning Architecture Mastery and Skill Development
Dates: September 14 – September 20
Internship: AI/ML Intern at SynerSense Pvt. Ltd.
Mentor: Praveen Kulkarni Sir
Focus
Building on the exploration phase from the previous week, this week was focused on honing practical skills in Deep Learning and Generative AI.
The objective was to go beyond theoretical understanding—by implementing, modifying, and analyzing various model architectures to strengthen applied proficiency.
This included working with CNNs, RNNs, LSTMs, Transformers, and experimenting with hybrid architectures for text-vision tasks. The week served as a self-driven training ground to refine both conceptual depth and hands-on fluency.
Goals for the Week
- Revisit core Deep Learning architectures and study their evolution
- Implement CNN, RNN, LSTM, and Transformer models using TensorFlow and PyTorch
- Compare model behaviors across different datasets and tasks
- Explore transfer learning and fine-tuning techniques on smaller datasets
- Document findings, insights, and reusable code patterns for future work
Tasks Completed
| Task | Status | Notes |
|---|---|---|
| Implemented CNN and RNN from scratch | ✅ Completed | Reinforced foundational understanding of feature extraction and sequence modeling |
| Fine-tuned pretrained models (ResNet, BERT) | ✅ Completed | Tested transfer learning for both vision and text domains |
| Compared model accuracy and convergence across architectures | ✅ Completed | Documented results with metrics and learning curves |
| Studied hybrid text-vision architectures | ✅ Completed | Analyzed ViLT and CLIP papers for multimodal understanding |
| Refactored earlier experiment code for modular use | ✅ Completed | Improved readability and reusability of scripts for future projects |
Key Learnings
- Architectural intuition matters. Understanding how layers interact helps in debugging and improving model performance.
- Transfer learning saves time and cost. Fine-tuning pretrained models delivers near-state-of-the-art results even with limited data.
- Visualization is crucial. Tools like TensorBoard and Grad-CAM provided deeper insight into model behavior.
- Hands-on iteration builds confidence. Implementing core components manually solidified practical understanding better than reading alone.
Challenges and Solutions
| Challenge | Solution |
|---|---|
| Long training times on large models | Used Google Colab Pro and smaller subsets of data for experimentation |
| Overfitting during fine-tuning | Applied data augmentation, dropout, and early stopping |
| Difficulty understanding multi-head attention outputs | Utilized visualization tools and Hugging Face tutorials for interpretability |
References
- Deep Learning with PyTorch – Official Tutorials
- TensorFlow Guide – Model Training and Transfer Learning
- The Illustrated Transformer (Jay Alammar)
- CLIP: Connecting Text and Vision (OpenAI Paper)
- ViLT: Vision-and-Language Transformer
Goals for Next Week
- Consolidate learning into mini-projects demonstrating CNN, RNN, and Transformer use cases
- Begin preparing final internship summary and presentation materials
- Document key insights into a technical report for the SynerSense knowledge base
Screenshots (Optional)
Screenshots of training curves, Grad-CAM visualizations, and model comparison charts showing accuracy and loss trends.
“Week 16 was about transforming understanding into ability—turning deep learning theory into practiced skill through experimentation, reflection, and iteration.”