123 points by johndoe 6 months ago flag hide 15 comments
john_doe 6 months ago next
This is really impressive! I wonder how well it scales with larger datasets.
another_user 6 months ago next
I believe the authors mentioned it still scales well, but I'd like to see some real-world tests before believing it fully.
john_doe 6 months ago next
From what I understand, yes. They've used ImageNet as a benchmark, and it outperformed existing models by around 2%.
ai_researcher 6 months ago prev next
This certainly looks like a promising development. Will be keen to understand the inner workings of the algorithm.
john_doe 6 months ago next
You might want to look at the arXiv paper they posted last week. It goes into feature attribution in quite some detail.
ai_researcher 6 months ago prev next
Thanks for the tip! I'll check it out.
original_poster 6 months ago prev next
Hi @data_scientist, it's indeed a different kind of optimisation algorithm. We're essentially using a graph-based approach for feature learning to combat overfitting.
john_doe 6 months ago prev next
I think they've done something with adapting backpropagation to allow convergence of loss at a higher rate.
ai_skeptic 6 months ago prev next
I'll wait for further confirmation from other researchers in the field before getting too excited.
original_poster 6 months ago next
I completely understand your skepticism, @ai_skeptic. Results will speak for themselves, and we welcome any criticism as we continue refining this process!