Very strong results on SSv2 and action anticipation, plus zero-shot robotics planning! And we also attached an LLM to the vision encoder and got strong numbers on PerceptionTest!

Check out the blog post (with link to paper and GitHub) above!

If you'd like to try yourself, the code has been added to our GitHub repository!

Qinco2 builds on Qinco with several optimizations, including beam search (increases accuracy+compute) and pre-selection (decreases compute). On the balance this leads to a more efficient method for similarity search.

The Qinco2 architecture builds on our previous Qinco work, which uses a neural network to implicitly parametrize code books for residual quantization. At each quantization step, a neural network is used in conjunction with the current vector to predict the next update.

Yes exactly.

Depending on how much you mutate it, keeping such libraries can also be very useful for reproducibility (which I didn't mention above).

But this can be difficult for other people that are trying to do something that doesn't fit in your framework (which happens often in research). There's always one or two things that simply don't fit. As a result, I'm finding myself writing more hacky/prototyping code these days.

In my PhD much of my code was hacky, and I think this set me back quite a bit. At some point I overcorrected towards building complex frameworks for my work, which let me try a lot of things (so long as I stayed within my own framework). This is more or less what you see in NeuralCompression.

Is there a particular reason this is considered an anti pattern? I'm actually curious.

I actually think there was quite a bit of spam here a month or two ago and it's already gotten better. Not sure if that's due to an effort on the part of the site admins or just basic engagement numbers shifting what gets into a feed.

Please make machine learners who still are children.

👋

I'm here!

Kinda has the y2k gaming energy (but without the CRT monitor)