The purpose of the project is to perform a Ratatouille Model Merging to obtain the best model to distinguish the different colorectal tissues and diseases.
This technique is very efficient because it allows to train multiple Neural Networks with the same architecture on different tasks separately (so called auxiliary tasks), fine-tune the models on the final task of interest, and then merge the models into a single definitive model.
In order to find the best model to classify the colorectal tissues, I decided to use auxiliary data that may help my model as much as possible, so I will train my models on datasets related to the human body: I will initially train the model on eye disease dataset and lungs disease dataset. After training the models on the data above, I will fine-tune them on the colorectal histology data, and successively merge the models.
The figure below (taken from the original paper and modified by me), explains better the procedure of the project:
The project has been done for the course "Neural Networks for Data Science and Applications" a/y 2023/2024