Diego Canez
banner
dgcnz.bsky.social
Diego Canez
@dgcnz.bsky.social
Self-supervised Learning, Efficient DL
University of Amsterdam
https://github.com/dgcnz
f mode collapse, all my homies hate mode collapse 😤👊
January 7, 2025 at 4:26 PM
Reposted by Diego Canez
Come and check out our poster on the UniReps workshop at @neuripsconf.bsky.social !

Thanks again to the amazing Giovanni Marchetti, Martina Scolamiero, and Danica Kragic ❤️

See more at: arxiv.org/abs/2409.10967
December 7, 2024 at 1:48 PM
Reposted by Diego Canez
🚨 PhD position alert! 🚨

I'm hiring a fully funded PhD student to work on mechanistic interpretability at @uva-amsterdam.bsky.social. If you're interested in reverse engineering modern deep learning architectures, please apply: vacatures.uva.nl/UvA/job/PhD-...
PhD Position in Mechanistic Interpretability
PhD Position in Mechanistic Interpretability
vacatures.uva.nl
December 2, 2024 at 7:36 PM
Reposted by Diego Canez
I am in Vancouver for NeurIPS 2024 until December 16th if you want to meet, DM or email me.
We have two accepted papers from my lab:
1. Building on Efficient Foundations: Effective Training of LLMs with Structured Feedforward Layers, on Wednesday, East Exhibit Hall A-C #2010 (1/3)
December 9, 2024 at 11:04 PM
Reposted by Diego Canez
Very happy to share our new paper that will appear in @unireps.bsky.social

We show improvements of zero-shot model stitching through invariance to symmetries in parameter space, and topological regularization of the latent spaces

more info below 👇
December 7, 2024 at 1:48 PM
First time being cited.

Ever since I started my Masters here at Amsterdam everything feels so surreal.

Back in Peru I had almost given up on the idea of doing research and now it scares me that it feels somewhat achievable.

I’m just so happy and thankful and aaaa

peace out and thanks (again) 🫡
December 5, 2024 at 4:13 PM
Reposted by Diego Canez
LoRA et al. enable personalised model generation and serving, which is crucial as finetuned models still outperform general ones in many tasks. However, serving a base model with many LoRAs is very inefficient! Now, there's a better way: enter Prompt Generation Networks, presented today #BMVC
November 26, 2024 at 7:28 AM
also try torch.compile, it’s free performance gains :)
now that people are paying attention again, here is your periodic reminder. Always run in bf16. always apply ROPE and attention softmax at float32 (as shown here)

github.com/xjdr-alt/ent...
November 24, 2024 at 9:29 PM
Reposted by Diego Canez
🚨 Happening today! 🚨

⏰ 4 PM CET / 10 AM EST
👩‍💻 Zoom link is in the link below
🏢 You can attend in person in L3.36, Science Park, University of Amsterdam

Join us for this exciting talk! :)
Soon, @erikjbekkers.bsky.social and @davidmknigge.bsky.social will give a talk elaborating even further on geometry-grounded representation learning in a NeurReps seminar. Make sure to mark the date! :)

⏰ November 21st, 4 PM CET
🔗 www.neurreps.org/speaker-seri...
November 21, 2024 at 9:07 AM
Reposted by Diego Canez
Soon, @erikjbekkers.bsky.social and @davidmknigge.bsky.social will give a talk elaborating even further on geometry-grounded representation learning in a NeurReps seminar. Make sure to mark the date! :)

⏰ November 21st, 4 PM CET
🔗 www.neurreps.org/speaker-seri...
November 19, 2024 at 4:14 PM
Reposted by Diego Canez
Yesterday, @erikjbekkers.bsky.social presented his vision on equivariance to IvI (at UvA), showcasing recent work on geometry-grounded representation learning - addressing fundamental limitations in geometric reasoning of current AI systems. 🤖
Exciting times ahead for geometric deep learning! 🌐 🤩
November 19, 2024 at 4:11 PM