This project introduces learners to the fundamental building blocks of deep learning using PyTorch. Learners begin by manipulating tensors and understanding real-world data representations. From there, they progressively build models, starting with linear regression and culminating in a simple multi-layer neural network (MLP). Using the Concrete Compressive Strength dataset, learners explore the relationship between model complexity, non-linearity, and prediction performance.

Building on the foundation laid in Project 1, this project guides learners through the core concepts and implementation of simple neural networks. Beginning with the perceptron algorithm, learners explore how neurons form networks, implement both forward and backward passes manually, and then use PyTorch to train and evaluate their models. The Heart Disease dataset provides practice with binary classification and helps learners understand the impact of model structure, non-linearity, and optimization.

This project introduces learners to multiclass classification using the Yeast dataset, a bioinformatics dataset involving 10 classes. Through four structured notebooks, learners explore new loss functions, optimization techniques, and evaluation metrics. They also investigate strategies for improving generalization, such as regularization, while tackling issues like class imbalance.

Building on multiclass classification, this project introduces deep neural networks using the CIFAR-10 image dataset. Learners explore the training challenges of deeper architectures, including initialization, optimization strategies, and stabilization techniques. The project concludes with simplified implementations of LeNet and AlexNet to transition into CNNs in the next project.

This project marks the transition from fully connected networks to convolutional architectures, equipping students with the core tools for computer vision tasks. Using the Oxford-IIIT Pet Dataset, learners explore how convolutional layers extract spatial patterns and textures from image data. Learners implement, train, and evaluate their first CNNs, and reflect on their performance relative to deep MLPs. The project concludes with a structural introduction to foundational CNNs like LeNet and VGG, preparing the ground for advanced vision architectures in the next module.

This project deepens learners' experience with convolutional architectures by exploring Transfer Learning and Advanced CNNs. Using the Plant Seedlings Classification dataset, learners work with real-world images of plants to build resilient models in the face of visual variability like lighting, angles, and background noise. The project focuses on modern strategies for improving training efficiency and model generalization, including fine-tuning pre-trained models and applying data augmentation techniques. This transition prepares learners to tackle production-ready image classification tasks using cutting-edge deep learning techniques.