Our workshop in numbers:
🖇️ 128 Papers
💬 8 Orals
🖋️ 564 Authors
✅ 40 Reviewers
🔊 7 Invited Speakers
👕 100 T-Shirts

🔥 Organizers: Paul Vicol, Mengye Ren, Renjie Liao, Naila Murray, Wei-Chiu Ma, Beidi Chen

#NeurIPS2024 #AdaptiveFoundationModels

This was work done during my internship with amazing folks @google @deep-mind.bsky.social: @nikparth1.bsky.social (joint-first) Ferjad Talfan @samuelalbanie.bsky.social Federico Yongqin Alessio & @olivierhenaff.bsky.social

Super excited about this direction of strong pretraining for smol models!

Bonus: Along the way, we found current state of CLIP zero-shot benchmarking in disarray—some test datasets have a seed std of ~12%!

We construct a stable & reliable set of evaluations (StableEval) inspired by the inverse-variance-weighting method, to prune out unreliable evals!

Finally, we scale all our insights to pretrain SoTA FLOP-efficient models across three different FLOP-scales: ACED-F{0,1,2}

Outperforming strong baselines including Apple's MobileCLIP, TinyCLIP and @datologyai.com CLIP models!

There's more! ACID and KD are complementary — they can be profitably combined, at scale! Our simple pretraining recipe ACED-ACIDistill showcases continued benefits as we scale to 26B samples seen!

We also show that ACID strongly outperforms KD across different reference/teacher training datasets, KD objectives, and student sizes.

Our ACID method shows very strong scaling properties as the size of the reference model increases, until we hit a saturation point — the optimal reference-student capacity ratio.

Further, ACID significantly outperforms KD as we scale up the reference/teacher sizes.

As our ACID method performs implicit distillation, we can further combine our data curation strategy with an explicit distillation objective, and conduct a series of experiments to determine the optimal combination strategy.

Our online curation method (ACID) uses large pretrained reference models (adopting from prior work: JEST) & we show a theoretical equivalence b/w KD and ACID (appx C in paper).

TLDR: We introduce an online data curation method that when coupled with simple softmax knowledge distillation produces a very effective pretraining recipe yielding SoTA inference-efficient two-tower contrastive VLMs!

Great paper! Why do you think it doesn’t make sense for pretraining to be made aware of the model being used in a few-shot setting downstream? Do you see any potential downsides of this kind of approach?

@bayesiankitten.bsky.social @dziadzio.bsky.social and I also work on continual pretraining :)

Congrats, super exciting!!