@babaktaati
Senior Scientist at the Kite Research Institute | Toronto Rehab - University Health Network. Associate Professor, Affiliated Scientist in the Department of Computer Science, University of Toronto (cross appointed at the Institute of Biomedical Engineering)
KITE Research Rounds today (Oct 27) at noon!
Speaker: Dr. Devin Brown, Professor of Neurology at the University of Michigan
Title: Post-stroke sleep apnea
kite-uhn.com/talk/researc...
Come by #ICCV2025 poster # 453 this morning to learn about our wok.
Project page: amirhossein-kz.github.io/lift/
CARE-PD is the largest publicly available archive of 3D mesh gait data for Parkinson’s Disease, collected across 9 cohorts from 8 clinical centers. It provides standardized, anonymized SMPL representations and benchmark protocols for clinical motion analysis on PD.
Our paper “CARE-PD: A Multi-Site Anonymized Clinical Dataset for Parkinson’s Disease Gait Assessment” has been accepted at #NeurIPS2025!
Project page: neurips2025.care-pd.ca
LIFT uses meta-learning to efficiently encode diverse signals (images, 3D shapes, and more) into multiscale latent representations for downstream tasks like generation and classification.
Paper: arxiv.org/abs/2503.15420
Project page: amirhossein-kz.github.io/lift/
If you're heading to #ICCV2025, come check out our work:
"LIFT: Latent Implicit Functions for Task- and Data-Agnostic Encoding"
🙋 I'll be at ICCV and interested in meeting
📜 Abstract: arxiv.org/abs/2506.10036
🌐 Project page: github.com/TaatiTeam/To...
Great collaboration with Javad Rajabi, Soroush Mehraban, Seyedmorteza Sadat.
TPG outperforms SDXL, PAG, and SEG in unconditional generation, and closely matches CFG in conditional tasks across all metrics.
TPG mirrors CFG by producing guidance vectors nearly orthogonal to ground-truth noise and maintaining strong guidance throughout denoising. Unlike PAG/SEG, it avoids negative alignment and weak updates, leading to more effective, high-quality generation.
TPG, like CFG, effectively recovers global structure and coarse details during early denoising steps. This stage is crucial for image quality and prompt alignment, as it establishes the foundation for structure and semantics.
In conditional generation, TPG produces better results and matches prompts more accurately than other attention-based methods like PAG and SEG. It also behaves more like CFG in how its guidance works and how often it applies.
TPG is a novel, training-free approach that boosts diffusion model performance by perturbing (shuffling) intermediate token representations. It improves both conditional and unconditional generation while also strengthening prompt alignment.
Our paper Token Perturbation Guidance (TPG) for Diffusion Models is accepted @neuripsconf.bsky.social 🥳
TPG is a simple & effective method based on token shuffling. It extends the benefits of CFG to broader settings, including unconditional generation.
arXiv: arxiv.org/abs/2506.10036
#NeurIPS2025
Developed with Dr. Nimish Mittal, Dr. Amol Deshpande & Andrea Sabo at Kite Research Institute | Toronto Rehab - UHN.
The HAT app is approved by Health Canada and, for now, is only available in Canada. We will release it in the US after receiving FDA approval.
The app guides users through 9 steps to return the Beighton Score, making hypermobility testing and EDS screening accessible to providers and the public.
Today we launched the HAT app during the Ehlers-Danlos Society's 2025 International Scientific Symposium, happening in Toronto this year.
kite-uhn.com/news/uhn-app...
Journal Club Sep 23: Arshak Rezvani and Dr. Maryam Mirian from UBC will present
DiffuseGaitNet: Diffusion Model Framework for Parkinson's Disease Gait Severity Assessment
Time: 13:00
Location: SRIC-11 Erin Room (W1160)
www.cs.toronto.edu/~taati/journ...
We did this back in 2023. I’d be curious to see how newer models perform if you get a chance to test them. The dataset is public. We had the first 15 seasons at the time, but I’m sure newer seasons are available now as well.
Today I had the idea to test LLMs on problems from the UK show Only Connect. Ofc, it turned out another group had already done it (@babaktaati.bsky.social)!
tl;dr: LLMs were pretty shit at the Connecting Wall and were often fooled by red herring distractors.
proceedings.neurips.cc/paper_files/...
I'm giving an IEEE Technical Talk at Queen's University, hosted by the Department of Electrical and Computer Engineering and Ingenuity Labs Research Institute, on May 15.
contentsharing.net/actions/emai...
Dr. Azadeh Yadollahi will be our next KITE Research Rounds Speaker
Title: Development of physiology-driven machine learning algorithms to assess cardio-respiratory function and diagnose sleep apnea
Date: Jan 27, 2025
Location: TRI 2nd floor auditorium
kite-uhn.com/talk/yadolla...
Journal Club Jan 21: We'll reviews two papers:
1) Scanning Trojaned Models Using Out-of-Distribution Samples (NeurIPS 2024), and
2) RODEO: Robust Outlier Detection via Exposing Adaptive Out-of-Distribution Samples (ICML 2024).
Journal Club Jan 8th: we read DiffuseMix: Label-Preserving Data Augmentation with Diffusion Models (CVPR 2024)
Date: Wednesday Jan 8
Time: 1 pm
Location: SRIC-11 Minkowski Room (W1148)
![The Dark Matter of AI [Mechanistic Interpretability]](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:t5c4ju4q6k34j6xysrhcoxmt/bafkreifee5measxm74wqrlt3ffnvqq5tmh66qstoynfgmtbiovrb2my3ly@jpeg)
www.youtube.com/watch?v=UGO_...
A super cool explanation of recent LLM mechanistic interpretability studies from anthropic etc. from Welch Labs (an awesome channel btw).
see also "The Dark Matter of Neural Networks?" by @colah.bsky.social transformer-circuits.pub/2024/july-up...
Scientist position at The Hospital for Sick Children in Toronto. We are looking for neuroscientists with interests in the molecular and cellular bases of cognition. The position is open-rank, and the successful candidate will also be appointed at U Toronto. Deadline is Feb 15th. Come join us!
Journal Club Dec 18: We will read "What Do Self-Supervised Vision Transformers Learn?" by Park et al. from ICLR 2023.
www.cs.toronto.edu/~taati/journ...
Journal Club Dec 11: We will review "Interfacing Foundation Models' Embeddings" from NeurIPS 2024.
www.cs.toronto.edu/~taati/journ...
Hi Christian, long time no see, indeed! Good to see you here. Thanks for letting me know about the link.