Michael Ekstrand

HPPC has certain advantages for writing super-fast code, too.

I'm more inclined towards HPPC since we have removed Fastutil from public-facing APIs. We can put together a two-way bridge between Java collections and HPPC collections for the public-facing APIs....

Fastutil wipes the floor with HPPC in terms of performance: http://java-performance.info/hashmap-overview-jdk-fastutil-goldman-sachs-hppc-koloboke-trove-january-2015/ I have re-run the current version of some of these benchmarks, and current FastUtil shows significant performance improvements over...

GS Collections has been rebranded to [Eclipse Collections](https://www.eclipse.org/collections/). I'm still interested in doing this.

So, front-to-back tests - yep, running over MovieLens is what we do. There are typically 2 versions of that: 1. A 'builld-serialize' test that makes sure we can train a...

> As for an accuracy test (accuracy would be the wrong word here BTW) I don't know that we have third party comparison for metric values of this algorithm on...

The build-serialize test will do 2 things: 1. End-to-end smoke test to make sure that training terminates & doesn't crash 2. Make sure the dependency structure allows models to be...

As we discussed this morning, the first steps should be to get a constant schedule working in the least-squares builder: 1. Create learning rate schedule interface 2. Create and test...

I have been investigating high CPU usage on low-power devices (e.g. Raspberry Pi Zero W), over in librespot-org/librespot#567; I suspect it may be this problem (or a related problem). At...

It looks like the raw metadata structure does not provide info to detect cloned files on macOS. I'm not sure what levels of cloning are supported on APFS, but on...