Magic Wand Image Classifier
Overview
This project was built for a Deep Learning course at BYU (Sept–Dec 2025). The goal was to design and train a deep neural network capable of real-time image classification on a custom dataset with structural similarities to MNIST: hand-drawn gesture images used as "magic wand" inputs.
This new system was to replace a live classification system being used on data captured from tracing an infrared LED on a real wand. The existing system used K-Nearest Neighbors (KNN) as its classifier, and while KNN is interpretable and easy to implement, its inference time scales linearly with dataset size, so the existing system was limited in dataset size in order to maintain acceptable speed, and thus sacrificed accuracy. The task was to replace it with a neural network that could classify faster without sacrificing accuracy.
The resulting model outperformed KNN on both metrics: lower latency at inference time and higher classification accuracy on the held-out test set. The project covered the full ML pipeline: data preprocessing, architecture design, training, and evaluation.
Technical Highlights
Deep Neural Network Architecture Design
Designed and trained a convolutional neural network architecture for image classification, iterating on layer depth, activation functions, and regularization to optimize for both speed and accuracy on the target dataset.
KNN Baseline Replacement
Replaced an existing K-Nearest Neighbors classifier with the neural network, achieving lower inference latency (constant-time vs. O(n) per query) while simultaneously improving classification accuracy.
Real-Time Classification Pipeline
The model was optimized for real-time inference, fast enough to classify gesture inputs with minimal perceptible delay, enabling interactive "magic wand" applications.