THthrowawaymath·Aug 20, 2018
Machine learning seems especially prone to attracting "aggregators" - resources whose sole purpose is to enthusiastically collect, promote and/or promulgate other resources for learning about the subject (or one of its sub-disciplines). I think we're starting to get at least one of these a week.
That observation aside, this isn't the first time this has been submitted by the author. That's not a problem in of itself, but I'd like to repeat part of my original critique [1] of this website when it was submitted a few months ago.
The author is pretty transparent about the way in which these resources are collected - they're automatically scraped from GitHub. I don't think that's intrinsically bad, but it does mean that due diligence is somewhat lacking. For example, consider the "model" page for Chess Reinforcement Learning: https://modelzoo
Read the rest
.co/model/chess-reinforcement-learning. That lists the README for the corresponding GitHub repository [2] verbatim. But the README specifically states that work on the model has been discontinued in favor of a new, better model's repository [3]. Additionally, the authors mention that they had difficulty with some of the implementation related to Self-Play.
Now I feel a philosophical point is in order. Ostensibly, an aggregator for resources should helpfully do all or most of the heavy lifting in 1) finding the resources, 2) vetting their completeness as advertised, and 3) maintaining up to date references for the resources it lists. There was no way to distinguish the limitations or obsolescence of the aforementioned model in first looking at it. Rather I had to open the page, click "Get Model" and read through the GitHub README on my own. I see substantially little difference between this activity and simply searching GitHub, considering that every single model I can find is listed on GitHub.
This is not to say that this project is bad, I just don't think it's particularly useful in its current form. I understand that's harsh feedback. But in my opinion the author would need to perform far more due diligence on the included models, the implementation artifacts of included models, the people implementing the models themselves and the methodology for model inclusion. As it stands this is presently a collection of models which are in varying states of completeness, developed by people who may or may not be authors of the papers they're implementing (or even have much implementation experience at all), and with varying levels of "out of the box" utility.
___________________
1. https://news.ycombinator.com/item?id=17311346
2. https://github.com/Zeta36/chess-alpha-zero
3. https://github.com/glinscott/leela-chess