On device personalization for text classification

[Marvin Moosbacher]

Hi all,

I know that currently there does not exist a full solution to perform on device personalization for text classification. But I would like to cover this topic in my master's thesis and refer to this experimental example for training on device with images: https://blog.tensorflow.org/2019/12/example-on-device-model-personalization.html

The main topic of my thesis is the added value of machine learning models on edge devices. My usecase is that I created a Tensorflow-Lite model for email classification (personal/impersonal) with the ModelMaker Library. Now I would like to describe how it could be solved to personalize the model for better results, especially when the dataset is not that good. Could you maybe give me a summary of the challenges you are working on for a complete solution, so that I can refer to them for presenting the current state of knowledge? I already tried to apply your presented technique for images to text classification for on device personalization but since I am not an real expert in that field it is not that easy for me, so I would like to refer on the challenges and possible solutions.

Thank you very much

Marvin