Today we’re joined by Babak Ehteshami Bejnordi (Babak Ehteshami Bejnordi), a Research Scientist at Qualcomm Follow us at @QCOMResearch, to discuss a few papers, including 'Conditional Channel Gated Networks for Task-Aware Continual Learning,' from last week's CVPR conference. twimlai.com/twiml-talk-385…

I was interviewed by The TWIML AI Podcast and we discussed our recent works at Qualcomm #AI Research on conditional computation using gated neural nets. Thank you Sam Charrington! Here is the link: twimlai.com/twiml-talk-385…

I just gave a tutorial on latent variable models with the focus on Variational Autoencoders and Normalizing Flows as two ways to avoid intractable inference. youtu.be/K9aIvuJhb7w This is part of the ourANU / Robotic Vision tutorial series on generative models.

Very excited to present "Natural Graph Networks" at NeurIPS next week. We use naturality - a generalisation of equivariance from category theory - to build an equivariant CNN that works on any graph. arxiv.org/abs/2007.08349 with Taco Cohen Max Welling

Be sure to take a look at great papers and demos from Qualcomm at #NeurIPS2020. Feel free to also tune in to #MLBites interview sessions with paper authors on Wed, 11:00-11:30; Thu, 8:40-9:10; Thu, 10:00-10:30 which I'll be hosting - available via neurips.cc/ExpoConference…

Do we need to process every single pixel in a video? TLDR: Compute the features only at the pixels that convey new information. Result: 4~5x less compute without any performance drop. Manuscript: arxiv.org/pdf/2104.11487…

How many frames are needed to reliably recognize an action? #FrameExit uses self-supervised gates to adjust the computation to the difficulty of the input video. Check out our #Oral #CVPR paper: arxiv.org/pdf/2104.13400… #CVPR2021 Great collaboration with ghodrati & Amir Habibian

Want a glimpse at the future of #AI? Check out our latest accepted papers and research breakthroughs at upcoming conferences, #ICLR and #CVPR2021. qualcomm.com/news/onq/2021/…

Really looking forward to this! We plan to release the lecture videos & course materials online

Qualcomm #AI Research is guided by purposeful innovation, passionate execution, and openness. Learn more: qualcomm.com/news/onq/2022/… #Qualcomm #machinelearning

dall-e 2 illustrations of my friends' twitter bios

Curious about what Qualcomm #AI Research has in store at #NeurIPS2022? Learn about our latest demos, papers, workshops, and other AI highlights: qualcomm.com/news/onq/2022/…

If you are at #NeurIPS2022, come visit us at the Qualcomm booth to check out our Expo Track demo: Conditional compute for on-device video understanding youtu.be/KuAx9xY235o via YouTube

The case for 4-bit precision: k-bit Inference Scaling Laws Shows that 4-bit precision is almost universally optimal for total model bits and zero-shot accuracy. arxiv.org/abs/2212.09720

We propose a dynamic tokenizer for ViTs, where the scale at which an image is processed varies based on the complexity of the image area. This means less computing for simple areas and more for complex, cluttered areas. Thanks to Amelie Royer, Jakob Havtorn, Tijmen Blankevoort

QIF Europe is an excellence award through which Qualcomm rewards and mentors the most innovative PhD students in Europe working on breakthrough #AI and #cybersecurity solutions. Congratulations Tycho van der Ouderaa Karsten Roth Siwei Zhang and Attri Bhattacharyya qualcomm.com/news/releases/…

🚀 Excited to share our latest work "Think Big, Generate Quick: LLM-to-SLM for Fast Autoregressive Decoding" now on arXiv! We're taking strides in making language models faster & more efficient on text generation tasks like translation & summarization.🔍 [arxiv.org/abs/2402.16844]

Proud to present our work on optimizing Mixture of Experts models for on-device generation speed: arxiv.org/pdf/2412.00099 We introduce a cache-aware routing that boosts memory efficiency of commonly used MoEs, improving generation throughput by 2×—all without retraining. Perfect

I'll be presenting CLUES🔍 at #NeurIPS2024 in person! Catch us at the poster session on: ⏰ Wed, Dec 11, 4:30–7:30 PM PST 📍 East Exhibit Hall A-C #1902 (Add it to your calendar: tinyurl.com/neurips-clues😊)

Our paper got a prize :) Cheers to lead author Johann Brehmer, and fellow co-authors Sönke Behrends, and Taco Cohen. Our results hint that yes, also at large scale of data and compute, if your data has symmetries, you might be better off building these into your network.