Adrian Bulat

Just a few days left (3rd of July) to submit your work on extreme quantization to the BiVision Workshop @ICCVConference We look forward to hear about you interesting work in this area! #ICCV2025 Call for papers: binarynetworks.io

Only a few days left (1st of august) to submit your papers to the Low-Bit Quantized Neural Networks (LBQNN) workshop @ICCVConference ⌛️

Join us at #iccv2023 at the workshop on Low-Bit Quantized Neural Networks (LBQNN). Learn about the latest developments in the field, foster collaborations or share your work by submitting a paper by 1st of august. Call for papers: sites.google.com/view/lbqnn-icc…

[Tech Blog] #CVPR2023 series ⑥ Introducing LASP, a text-to-text optimization method for vision language adaptation that sets a new state-of-the-art for #fewshot and #zeroshot image classification. #AI #VisionLanguage research.samsung.com/blog/LASP-Text…

📢📢📢📢Check our new Arxiv paper: Variational Prompt Tuning Improves Generalization of Vision-Language Models Link: arxiv.org/abs/2210.02390 w/ Enrique Sanchez, @AdrianBulat, @victorturrisi, @cgmsnoek, Georgios Tzimiropoulos, Brais Martinez at @samsungresearch 🧵1/5

Our work "Bit-Mixer: Mixed-precision networks with runtime bit-width selection" has being accepted at #ICCV2021 @ICCV_2021 ! Check it out if want to learn more about how you can quantize your model's blocks to different bit-widths at runtime. Preprint: arxiv.org/abs/2103.17267

Code will be released soon, in the meantime you can check our mini-framework used to train the models: github.com/1adrianb/binar… n/n

2) Dynamically selecting at runtime a single expert (i.e. weight) can boost the accuracy of Binary Networks at little additional computational cost. 3) Multi-staged approaches and feature-level distillation is necessary for high accuracy. 2/n

If you want to find more about getting high performing binary networks using conditional computing pass by our poster session tomorrow at @iclr_conf arxiv: arxiv.org/abs/2010.03558 1) Wider is better than deeper for BNNs, especially when combined with grouped convolutions. 1/n

Journal club today at 1: We'll read "Pre-training strategies and datasets for facial representation learning" by @AdrianBulat et al. @LeifSimmatis will be leading the discussion. @KITETrainees @AGEWELL_NCE @TRI_UHN cs.toronto.edu/~taati/journal…

In the past I've thought that having a comment system on arXiv would enable productive discussion and identification of errors in papers which were previously only known to a select few. Based on #ICLR2021's OpenReview, I think it would just be people saying "plz cite me"

Excited to announce the Workshop on Binary Networks held in conjunction with #CVPR2021. Great place for learning about efficient deep learning models from our amazing invited speakers. Paper submission will open soon. For more details, please check: binarynetworks.io

Interested in automatic #emotion estimation from faces in natural conditions? 😁☹️😠😧😮 Checkout our @NatMachIntell paper - we estimate discrete emotion, continuous valence & arousal levels and facial landmarks w/ very high accuracy! ⬇️ youtube.com/watch?v=J8Skph… @SpringerNature

Videos of #EECVC2020 are available! Day 2: Dr. Daniel Barath talk "State of the art in robust geometry estimation" and Dr. @AdrianBulat "Efficient deep models via network binarization/quantization" youtu.be/8LJcaL2Nz-w

Videos of #EECVC2020 are available! Day 1: Dr. @amy_tabb talk "Computer vision and robotics in agriculture" and Dr. @ftmguney talk "Optical flow" youtu.be/jC3ESFq32TQ

@ducha_aiki Totally agree! Best email is the one you can actually access. Same happened to me, my university email got terminated on the graduation day. Needless to say that now I have multiple accounts on pretty much all academic websites and get invites or spam on all of them :)

At the beginning of my career, I was told that using personal (e.g. gmail) email account is not professional. That is rubbish. Why? Because institutional email addresses rot quickly. E.g., today I got a message: "I have graduated from XXX and this email will be canceled soon"

@JeanKossaifi @JmlrOrg @zacharylipton @arinbjornkol @arankhanna @furlanel @AnimaAnandkumar Congratulations! Great work and an interesting paper to read!

Tensor Regression Networks is published @JmlrOrg! With @zacharylipton @arinbjornkol @arankhanna @furlanel @AnimaAnandkumar

Today is the 2nd day of the #EECVC2020. Register now, if you haven't already, to listen to our speakers at 19:00 GMT+3 Adrian Bulat about CNN quantization and Dr. Daniel Barath about modern RANSACs! eecvc.com/register-now-f…

@JeanKossaifi @nvidia @AnimaAnandkumar @NvidiaAI Congratulation! May the tensors be with you, and the luck too! :)

Really excited to be joining @nvidia as a senior research scientist, working in #ML and #tensor methods with @AnimaAnandkumar and the wonderful team there @NvidiaAI

The #NeurIPS reviewer option “I will consider not reviewing for NeurIPS again if this submission is accepted.” is ridiculous. It encourages a toxic attitude towards reviewing, and when it is actually used, can be incredibly hurtful to our young scholars. @NeurIPSConf