Berk Tınaz

𝐄𝐱𝐜𝐢𝐭𝐞𝐝 𝐭𝐨 𝐬𝐡𝐚𝐫𝐞 𝐭𝐡𝐞 𝐩𝐫𝐞𝐥𝐢𝐦𝐢𝐧𝐚𝐫𝐲 𝐫𝐞𝐥𝐞𝐚𝐬𝐞 𝐨𝐟 𝐌𝐨𝐬𝐚𝐢𝐜𝐌𝐑𝐈. To our knowledge, it is the 𝐥𝐚𝐫𝐠𝐞𝐬𝐭 𝐚𝐧𝐝 𝐦𝐨𝐬𝐭 𝐝𝐢𝐯𝐞𝐫𝐬𝐞 𝐨𝐩𝐞𝐧-𝐬𝐨𝐮𝐫𝐜𝐞 𝐫𝐚𝐰 𝐦𝐮𝐬𝐜𝐮𝐥𝐨𝐬𝐤𝐞𝐥𝐞𝐭𝐚𝐥 𝐌𝐑𝐈 𝐝𝐚𝐭𝐚𝐬𝐞𝐭 to date: 2,671 volumes, 80,156 slices, 454 patients, and 10 anatomies, with broad variation in contrast, orientation, and coil configuration. This is a preliminary release and reflects only a fraction of the broader dataset we have been curating. We plan to expand MosaicMRI across institutions, field strengths, and new clinical challenges, with the goal of supporting both core 𝐌𝐑𝐈 𝐩𝐫𝐨𝐛𝐥𝐞𝐦𝐬 (accelerated reconstruction, low-field MRI, motion compensation) and broader 𝐀𝐈 𝐚𝐧𝐝 𝐟𝐨𝐮𝐧𝐝𝐚𝐭𝐢𝐨𝐧-𝐦𝐨𝐝𝐞𝐥 𝐪𝐮𝐞𝐬𝐭𝐢𝐨𝐧𝐬 in science and healthcare, including robustness, generalization, scaling laws, heterogeneous data mixtures, continual learning, and test-time adaptation. 𝐖𝐞𝐛𝐬𝐢𝐭𝐞: mosaicmri.ai 𝐏𝐚𝐩𝐞𝐫: arxiv.org/abs/2604.11762 𝐂𝐨𝐝𝐞: github.com/AIF4S/mosaicmri 𝐀𝐜𝐜𝐞𝐬𝐬: #access" target="_blank" rel="nofollow noopener">mosaicmri.ai/#access 𝐁𝐞𝐧𝐜𝐡𝐦𝐚𝐫𝐤: mosaicmri.ai/benchmark/ 𝐄𝐚𝐫𝐥𝐲 𝐞𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧 𝐭𝐫𝐚𝐜𝐤𝐬 𝐚𝐫𝐞 𝐧𝐨𝐰 𝐥𝐢𝐯𝐞. Official benchmark tracks and community challenges will follow. Grateful to NIH National Library of Medicine (NLM) for supporting this project.

Our analysis reveals how interpretable concepts emerge and evolve across diffusion timesteps--and how this structure can be leveraged for targeted interventions. 🔗 Workshop: openreview.net/pdf?id=plCImt7… 🔗 arXiv: arxiv.org/pdf/2504.15473 With: @zalan_fabian @mahdisoltanol (8/n,n=8)

We then explore style control. Edits in early steps primarily influence global composition, while edits in mid-stage timesteps affect artistic features--e.g., a cartoon-like rendering from feature #1314 or a grainy texture from #524. (7/n)

I’ll be at #CVPR2025 presenting a spotlight talk at the VisCon Workshop on our latest work: "Emergence and Evolution of Interpretable Concepts in Diffusion Models" 🕒 June 12th, 3:15PM CDT A short thread summarizing the paper 🧵

@akyurekekin @afeyzaakyurek @OpenAI Congrats, Ekin!

✨ Big life updates ✨ - @afeyzaakyurek and I welcomed our baby! - Successfully defended my PhD and graduated from MIT 🎓 - Joined @OpenAI 🍓 Excited for what's next!

For more details about severity encoding and more results, check out our paper: arxiv.org/pdf/2309.06642  (@zalan_fabian @mahdisoltanol) Stay tuned for the GitHub and project page! (7/n, n=7)

Moreover, Flash acts as a wrapper that can enhance any latent diffusion-based solver, making it sample-adaptive and faster (up to 10x), while performing better in terms of both perceptual (LPIPS, FID) and distortion (PSNR, SSIM) metrics. A snippet from our results table: (6/n)

🚀”Adapt and Diffuse: Sample-Adaptive Reconstruction via Latent Diffusion Models” will appear in @icmlconf as a Spotlight! TLDR: Our method adjusts the number of diffusion steps based on problem difficulty, significantly saving on compute while improving quality! (1/n)

We are looking for Summer/Fall grad interns as part of the USC Center on AI Foundations for the Sciences (AIF4S)(sites.usc.edu/aif4s/). Deadline is March 12 detail and form here sites.usc.edu/aif4s/openings/.