A series of core deep learning projects covering mathematical optimization, computer vision, sequence modeling, and representation learning.
Quick Facts
- Context: CS583 Deep Learning Coursework (Spring 2026)
- Tech Stack: Python, TensorFlow/Keras, Keras Tuner, NumPy, scikit-learn, Matplotlib
- Links: None available
Overview and Problem
This portfolio encompasses four major assignments designed to build intuition and practical skills in deep learning. The projects span from from-scratch optimization algorithms to complex sequence models and autoencoders.
What I Built
- Implemented Logistic Regression and multiple optimizers (Batch GD, SGD, Mini-Batch GD) from scratch using NumPy.
- Designed and trained a custom Convolutional Neural Network (CNN) architecture with Batch Normalization for CIFAR-10 image classification.
- Built a character-level Encoder-Decoder seq2seq model using Bidirectional LSTMs for English-to-Spanish translation.
- Engineered a dense autoencoder with a 2D bottleneck layer, eventually extending it to a supervised autoencoder with a joint reconstruction and classification objective.
Key Results and Impact
- Achieved 97.37% test accuracy on Breast Cancer classification using SGD, and demonstrated the generalization effect of L2 regularization.
- Attained 82.02% test accuracy on CIFAR-10 after addressing overfitting and conducting systematic hyperparameter tuning.
- Yielded >90% classification accuracy on MNIST using supervised 2D latent features, confirming the value of label supervision for discriminative representations.
Related: Projects MOC