Possible fault in your evaluation process about torchdiffeq HOT 3 CLOSED

Dear Yuval,

I think you've identified a real weakness in that experiment, which is that we used an unnecessarily large baseline network on this simple task. We chose the baselines architecture to be close to standard resnets, but we should have found the smallest such architecture that still had good performance.

The other main points of that experiment, which is that the memory usage is O(1) and that ODEnets can indeed be trained using the adjoint sensitivity method, I think still stand. But in light of your result, I think you're right that our table doesn't demonstrate that ODEs can achieve the same performance with fewer parameters than the best resnet architecture. In fact, we could also have tried parameterizing a resnet to have parameters tied across layers, output by a hypernetwork as well.

Many thanks for diplomatically pointing out this issue! We'll discuss how to update the paper.

from torchdiffeq.

It took me way too long to get to this, but I finally updated the paper to simply remove the unsupported claim about parameter efficiency, and added @YuvalFrommer to the acknowledgments:

https://arxiv.org/pdf/1806.07366v5.pdf

Thanks again!

from torchdiffeq.

from torchdiffeq.