This repository presents a journey through three progressive milestones of the "Basics of Deep Learning" course. Each milestone explores different modeling approaches, frameworks, and levels of abstraction - from scratch implementations with NumPy to full-scale CNNs in PyTorch. The final milestone dives into a fine-grained visual classification task using the Stanford Cars dataset.
Each milestone has its own dedicated folder and README.md that includes full documentation, architectures, training processes, evaluation results, and conclusion
We began by building a feedforward neural network entirely from scratch using NumPy to classify between two hand gestures (digits 6 and 8) from the Sign Language Digits Dataset. The project involved:
- Manual implementation of forward and backward propagation.
- Use of the Sigmoid activation function and binary cross-entropy loss.
- Mini-batch gradient descent and clean performance visualization.
The model reached an impressive 95.6% accuracy and taught us foundational deep learning concepts from the ground up.
In this stage, we transitioned to PyTorch, re-implementing the binary classifier and expanding it into a multi-class classification model capable of recognizing all ten digits (0–9). Highlights include:
- Structured training pipeline using DataLoader, GPU acceleration, and autograd.
- Development of deeper architectures with dropout and ReLU.
- Hyperparameter tuning, early stopping, and custom Dataset class.
The multiclass model achieved 98.93% validation accuracy and 100% test accuracy on unseen .npy examples.
The final milestone tackled a fine-grained classification task using the Stanford Cars Dataset with 196 car categories. We explored three configurations:
- Fine-tuned pre-trained models on ImageNet.
- Best model: ResNet-50 fine-tuned with augmentation and dropout, achieving 73.78% accuracy.
- Extracted feature embeddings from CNNs.
- Performed top-k retrieval with Euclidean and Cosine distance.
- Best setup: ResNet-50 + KNN (k=10), reaching 76.77% accuracy.
- Custom architecture with 4 conv blocks, batchnorm, and augmentation.
- Final model reached 72.26% accuracy, demonstrating competitive results even without pre-training.
Through these milestones, we experienced the evolution of deep learning practices:
- From scratch to frameworks - We saw how using NumPy teaches the mechanics, while PyTorch boosts flexibility and power.
- From binary to multiclass and fine-grained tasks - We advanced from simple digit recognition to subtle distinctions between car models.
- From classification to retrieval - We broadened our thinking from prediction to semantic similarity.
This portfolio not only enhanced our technical skills but also strengthened our understanding of model selection, architecture design, performance evaluation, and deployment considerations.
 Colab |
 Overleaf |
 Kaggle |
 YOLOv8 |
 Python |
 NumPy |
 Pandas |
 Matplotlib |
PyTorch |
 scikit-learn |