Danilo de Goede

Working on Adaptable Foundation Models | PhD student @UvA_Amsterdam

[Danilo de Goede @danilodegoede]

@j0hngou Thanks❤️

[Danilo de Goede @danilodegoede]

How many tokens does it take to capture the gist of an image?

Introducing GIST: an image tokenizer that learns to preserve what matters and chooses how many tokens each image needs, in one forward pass.

Now on arXiv 🎉🎉
https://t.co/M8rVDOTPtx

Demo: https://t.co/EcUkeg9eah

[Danilo de Goede @danilodegoede]

@mmderakhshani @y_m_asano @cgmsnoek

[Danilo de Goede @danilodegoede]

Compression is a choice about what to forget.

Try GIST's two controls in the interactive demo:
https://t.co/EcUkeg9eah

Paper: https://t.co/M8rVDOTPtx

With Mohammad Mahdi Derakhshani, Yuki M. Asano & Cees G. M. Snoek.

[Walter Simoncini @iWSimoncini]

🪗 Your Transformer Is Secretly Elastic 🧵👇

Transformer families come in a few fixed sizes, so you commit to one for your hardware and latency constraints, even as query loads vary. What if a single pre-trained model could serve every budget?

We learn one importance ranking over every MLP neuron and attention head, once, with frozen weights. At inference time, the model can then be pruned to any computational budget by dropping its least important units, without retraining.

Across 27 vision and text encoders spanning 22M to 8B parameters, performance degrades smoothly up to 60% sparsity. For example, pruning DINOv3 ViT-H+/16 to 50% sparsity removes 427M parameters, for only a 0.8% drop in ImageNet-1k linear probing.

Super fun project in collaboration with Dheeraj Varghese (@dhevarghese), supervised by Gertjan J. Burghouts (@gjburghouts) and Cees Snoek (@cgmsnoek)

🌐 Project page: https://t.co/zBNxee1Xyh