<aside> ℹ️ What is this?
Frontiers is a learning project focused on recent trends in machine learning (i.e. computational alchemy) that is run by Alex Amirejibi and Paul Bricman. It consists of reading and replicating a range of curated papers in the field.
Where do the papers come from?
We are using the Ought Machine Learning Reading List as a starting point, as it provides an expert-curated list of papers that span a number of topics, from scaling laws to uncertainty quantification. Conveniently, it is also organized in terms of levels, making more sophisticated papers accessible by virtue of the ones leading up to them.
</aside>
<aside> ℹ️ Don’t these papers require astronomical amounts of compute to replicate?
Some of them certainly do, so we triage them in terms of their reproducibility potential according to the following categories:
</aside>
tractable: we might be able to reproduce the entirety of the paperpartially tractable: we might be able to reproduce a significant part of the paper (e.g. runs with lower model sizes)intractable: it’s largely out of reach (e.g. authors hire a whole cohort of contractors for labeling)unapplicable: the paper rather documents a survey or a conceptual discussion