Shady Abu-Hussein (@ShadyAbh) profile picture

Postdoctoral Researcher at the University of Cambridge (Previously: Samsung Research, Intel, IBM Research).

Joined December 2021
Photo by Shady Abu-Hussein (@ShadyAbh)Photo by Shady Abu-Hussein (@ShadyAbh)Photo by Shady Abu-Hussein (@ShadyAbh)Photo by Shady Abu-Hussein (@ShadyAbh)Photo by Shady Abu-Hussein (@ShadyAbh)Photo by Shady Abu-Hussein (@ShadyAbh)Photo by Shady Abu-Hussein (@ShadyAbh)
Shady Abu-Hussein retweeted
Introducing interpolated FID: an variant of reconstruction FID that is highly correlated to diffusion generation FID. It is well known that VAE rFID is not correlated with diffusion gFID. We propose a simple variant that achieves correlation 0.9! huggingface.co/papers/2603.0…
4
2
6
813
Shady Abu-Hussein retweeted
DeciMamba, the first context extension method for Mamba, is accepted to #ICLR2025! πŸŽ‰ New revision with more long-context results: arxiv.org/abs/2406.14528 github.com/assafbk/DeciMamba Special thanks to @ItamarZimerman @ShadyAbh @nadavcohen @amirgloberson @liorwolf @RGiryes !
New Work! 🐍 What prevents Mamba from extrapolating to sequences that are significantly longer than those it was trained on? Furthermore, can Mamba solve long-range NLP tasks using short-range training only? 🧡🧡🧡
1
7
2
24
3,465
Shady Abu-Hussein retweeted
Excited to announce that our paper on efficient diffusion models has been accepted to NeurIPS! Huge kudos to @ShadyAbh for the incredible work. Developing carefully the math and implementing it to achieve strong results took a lot of effort, and it great to see it pay off!
(1/6) 🌟 New Paper! 🌟 Is it possible to sample images from a diffusion model using less than one step? Answer: Yes! ➑️ We present UDPM, a novel approach to image generation that outperforms existing models in efficiency and quality. πŸš€πŸ“Š arxiv.org/abs/2305.16269
1
2
35
1,983
(1/6) 🌟 New Paper! 🌟 Is it possible to sample images from a diffusion model using less than one step? Answer: Yes! ➑️ We present UDPM, a novel approach to image generation that outperforms existing models in efficiency and quality. πŸš€πŸ“Š arxiv.org/abs/2305.16269
2
11
2
63
6,702
(5/6) * UDPM has an interpretable and interpolable latent space, making it superior to traditional DDPMs in various applications. πŸŒπŸ” * UDPM has smaller latent variables, hence one may modify the noise maps of each step and study its effect on the generated image:
1
3
201
Shady Abu-Hussein retweeted
New Work! 🐍 What prevents Mamba from extrapolating to sequences that are significantly longer than those it was trained on? Furthermore, can Mamba solve long-range NLP tasks using short-range training only? 🧡🧡🧡
1
27
5
122
26,584
UDPM: Upsampling Diffusion Probabilistic Models. arxiv.org/abs/2305.16269
1
50
Shady Abu-Hussein retweeted
I like the idea behind "Upsampling Diffusion Probabilistic Models" (arxiv.org/pdf/2305.16269.pdf). Instead using a fixed sized noised sample, use a smaller sized sample for noiser timesteps
2
15
58
10,607