use case: transfer learning / finetuning

[gfrogat]

Most people don't train models from scratch but fine-tune a pretrained model instead.

The resulting dynamics of the model might therefore be different. On a different note: is not completely clear whether the observed oscillations are due to changes in the convnet filters or in the fully connected parts of the network

[gfrogat]

In this context it might be interesting to compare the learning / fine tuning of an image classifier on summer and winter images.

[thinkh]

We would need a dataset with summer and winter images for this use case.

[gfrogat]

we can illustrate similar effects by experimenting on different random folds, i.e. cross validation.