This project was intended to test the limits of the ViT on a tough dementia dataset. The data used can be found on HuggingFace at: https://huggingface.co/datasets/Falah/Alzheimer_MRI. The dataset has 4 classes:
- 0: mild_demented
- 1: moderate_demented
- 2: non_demented
- 3: very_mild_demented
The project follows closely the following tutorials:
- https://www.youtube.com/watch?v=r88L_yLJ4CE&ab_channel=code_your_own_AI
- https://www.youtube.com/watch?v=qU7wO02urYU&ab_channel=JamesBriggs
I modify the code presented in the video and tune all parameters to optimize performance using mostly the same libraries and tools. This is a practice project for myself as I return to coding/designing ML models after dedicating time to AI/ML theory (model architectures, transfer learning)
After hyperparameter tuning, the highest testing accuracy achieved is 66% using a learning rate of 2e-5, a weight decay of 0.001, batch size of 8, and iterating over 5 epochs.
