Yes from what I'm seeing you definitely need to use a neural net.

Where is that indicated? Because I suspect most of the top LB (if not all) are using boosting models

Yes I also am concerned that to be in the top LB, we have to use boosting models which is not mentioned in the info of this challenge.

Your solution must include:

A centralised model, trained using the combined dataset.

A federated learning approach, trained using the split datasets.

The aim is to prove that MPEG-G files are more efficient due to less data movement and better compression, and that federated learning preserves performance across sites better than a centralised model. You must:

Define Clients: Treat each folder (or grouping by participants/site) as a "location".

Create a Federated Dataset Loader: Each client should have its own dataset split (e.g., location1_data, location2_data, ...) and convert each into a PyTorch Dataset.

Define the Same PyTorch Model: Same architecture as centralized model.

Implement FL with Flower or PySyft

Train the FL Model

Make a submission to Zindi

In your final solution submission , you will need to output three submission files, one from your centralised model, one from your federated learning model and a technical explanation of the innovation and approach used. You will also need to indicate training time, inference time, resources used and any observations on .mgb vs uncompressed formats.

But we can also apply Federated learning to xgboost for example (although not sure about the performance). So why the restriction

Can't tell, let's confirm from the organizers😅.

If you are getting this score: 3.16e-7 without post processing you should definitely submit it as it is valid

But the thing is, it's a DT based model. Federated learning using that is complex.

Also the competition doesnot want us to use any other model. It mentioned to use pytorch. That's why so much confusion. :)

federated learning is a must! but I have seen an xgboost federated learning implementation. You can try that. Also for that score what is your local cv?

Boosting models are ready, as well as the NN. But the performance between these two are quite large. No one is yet sure how the final evaluation will be done during code checking.

Will those codes having boosting models be accepted as a valid one?

I think they will focus on both federated and centralized approaches and then carbon emission, basically how long it takes to train.

For now, we need to hear from the organizers on whether they prefer a pytorch model or any. I would however urge you to follow the rules.

@Amy_Bray please clarify this for us.

Unless clarified, we are unable to understand whether to move forward with neural networks or other models.

Exactly!