This week, I fully implemented the Practice page within the existing React/Vite localhost framework. Currently, the practice mode cycles through all 26 ASL letter signs (A–Z), evaluating each response as correct or incorrect. For testing purposes, a 1-in-4 probability of an incorrect answer is simulated. Upon completing all 26 characters, the user is presented with a results summary screen displaying their overall performance.
For the machine learning subsystem, I built and validated the data generation and training pipeline The pipeline currently generates a synthetic dataset and feeds it through a neural network classifier. I initially had an issue where the output layer was hardcoded to 5 classes while the dataset contained 36, causing the model to perform at random chance with ~5% accuracy. After fixing the output layer to dynamically match the number of classes in the dataset, the model achieved 100% accuracy on the synthetic data. This is expected as the synthetic data is generated from a distinct random base vector with minimal noise. This result confirms that the pipeline is functioning correctly and is ready to be evaluated against real sensor data.
Next week, I plan to design and implement the transcription mode. I will also begin forming the actual training dataset from collected data points. Once an initial set of real samples is gathered, I will apply data augmentation techniques to artificially expand the training set. This is a standard approach for sensor datasets where collection time is limited and will allow us to increase dataset size significantly without requiring more physical recording sessions. If time permits, I will also set up a REST API endpoint that exposes the model’s predictions to the frontend.