Fixed training vs. Fine tuning - start with Fixed Training

[epowell101]

As a first step we can try a fixed tuning, using the existing weights, plus a classifier that we train with our labels.

The approach suggested is to use a MLP for classification where the Bert4eth model is used for the embedding.

[epowell101]

Note though that they were unable to fine-tune Bert4eth for the de-anonymization task as they only had 288 pairs for the ENS dataset and 182 pairs for tornado.