Parametric Matrix Models

arxiv.org

66 points·evanb··7 comments

Looks like a cool idea and it could benefit from a more complete and detailed illustration and explanation of the architecture.

Reads like the authors skip to implications before clarifying the design.

Also, a stylistic sidenote, narrower columns of text are much easier to read, newspapers and journals do this for good reason

First author here. I agree with your points, we were constrained by the format and (especially) writing style expected by the journal we're submitting to. The Methods sections contain more explicit explanations as well as other analyses.

Huh, interesting.

According to the authors, these "parameteric matrix models" or PMMs outperform:

* commonly used (zero- or low-parameter) regression models like XGBoost, random forests, kNN, and support vector machines on a variety of regression tasks, and

* DNNs with 10x to 100x more parameters on small-scale image classification tasks like MNIST variants, CIFAR-10, and CIFAR-100 -- albeit with a lot of feature engineering.

It looks promising, but I cannot find a link to the authors' code for replicating their experiments.

Morten was my PhD adviser. I'll ask him what's up.

All of the code and data will be released with the peer-reviewed published version. If I remember, I'll come back to this thread and post the link.

Thank you for taking the time to update everyone on HN.

I've added your work to my reading list.

Of course - I'm glad to see people are interested. I look forward to any feedback on our work either public or private.

Reply on news.ycombinator.com