Sara Atito Ali Ahmed

[Sara Atito Ali Ahmed @atito_sara]

@The_Mardini أنا مش بشتري وابيع غير ذهب و Nasdaq و UK100
ممكن ترشح حجات تاني ممكن تبقى كويسه للمضاربه؟

[Sara Atito Ali Ahmed @atito_sara]

@The_Mardini اكتر حد بمشي وراه وكل مره بتطلع صح 🙌🙌🙌🙌 u r the best

[Sara Atito Ali Ahmed @atito_sara]

@The_Mardini ان شاء الله هنكسببب

[Sara Atito Ali Ahmed @atito_sara]

@The_Mardini 💃💃💃💃💃

[Sara Atito Ali Ahmed @atito_sara]

@The_Mardini اي تصريحات؟ فين احسن مكان أتابع منه الأخبار، بقى في fake news كتير جدا دلوقتي

[Sara Atito Ali Ahmed @atito_sara]

@xxxnell @_akhaliq Thanks Namuk for the great analysis. Please have a look at my paper “SiT: self-supervised vision transformers” one of the early papers that leverage self-supervised learning for vision transformers, proposing MIM with vision transformers and combining it with contrastive learning

[Sara Atito Ali Ahmed @atito_sara]

@abursuc @david_picard It exists 😂 GMML is all you need which is group masked model learning is all you need 😅

[Sara Atito Ali Ahmed @atito_sara]

@jd92wang @NeurIPSConf Same here, we even tried to further initiate the discussion but no response!

[Sara Atito Ali Ahmed @atito_sara]

Our spotlight talk on “Towards Multi-concept Self-supervised Learning” #transformers4vision #cvpr2022 #CVSSP @cvssp_research full spotlight > https://t.co/JAyA6Jq5He

[CVSSP Research @cvssp_research]

Our startup #SensusFuturis is at the Farnborough Security and Policing Exhibition (pod 16&17). Explore #AI solutions for #FaceRecognition, #ImageSearch, #VideoAnalytics, #Identity & #Privacy. https://t.co/dkG7kC7zKL   #securityandpolicing

[Sara Atito Ali Ahmed @atito_sara]

@ylecun @lexfridman True, but we managed to make it work with high performance on small datasets. Unfortunately the paper got rejected because we didn’t train on ImageNet which is not feasible for us as a small group.

[Kosta Derpanis @CSProfKGD]

#KostasThoughts #CVPR2022: There is no shame getting the initial review wrong, e.g., missing an important detail. It is wrong when you let your ego get into the way of arriving to the correct decision.

[Sara Atito Ali Ahmed @atito_sara]

@y_m_asano You are right. But the GMML is not just for augmentation, but rather extract the most information from the available data which made it feasible to obtain high accuracy on small datasets when trained from scratch. I think even with more data, GMML will be beneficial.

[Sara Atito Ali Ahmed @atito_sara]

@y_m_asano and this masking encourages the network to learn from semantic information present in the locality of the masked tokens to recover the missing information as well as the uncorrupted parts, and hence, the network will unlikely to overfit.
2/2

[Sara Atito Ali Ahmed @atito_sara]

@y_m_asano I think the situation is different in vision, but also it depends on how the network is trained. Unlike NLP, in SiT, we apply severe corruption on a [group] of connected patches
1/2

[Sara Atito Ali Ahmed @atito_sara]

@y_m_asano You can check “SiT: self supervised vision transformer”. It’s the same idea, proposed long ago, and because we don’t have enough compute, we only show the results on small datasets where we outperform the SOTA at that time.

[Sara Atito Ali Ahmed @atito_sara]

@francoisfleuret @ak92501 @Meta You are right. I was really disappointed at that time, and to be honest, I did not expect that anyone will pay attention to my respond as I am not even a twitter user.

[#CVPR2027 @CVPR]

Official response from our #CVPR2022 program chairs.

[Sara Atito Ali Ahmed @atito_sara]

@JFPuget @ak92501 @Meta https://t.co/wB5h1r9x5A