Associate Professor, Electrical & Computer Engineering @TMU_ecbe Faculty Affiliate @VectorInst
Ontario, Canada
Joined November 2017
- Tweets646
- Following76
- Followers139
- Likes979
Reza Samavi retweeted
Superposition as Lossy Compression — Measure with Sparse Autoencoders and Connect to Adversarial ...
Leonard Bereska, Zoe Tzifa-Kratira, Reza Samavi, Stratis Gavves.
Action editor: Pin-Yu Chen.
openreview.net/forum?id=qaNP…
#autoencoder #superposition #ne
Reza Samavi retweeted
VLMs (GPT-4o, Gemini, Qwen-VL, LLaVA…) look impressive — until you shift an image by 1 pixel.
A tiny, meaning-preserving change → a completely different answer.
This isn’t adversarial — it’s natural variation.
Watch 👇
Reza Samavi retweeted
Check out the amazing @let4all Mentorship Workshop happening in just two weeks!
Theme: Harnessing AI for Research, Learning, and Communicating
Featuring an amazing lineup of speakers and also hands-on experience with ChatGPT Pro to see how to use AI to help your research.
Announcing the 7th Learning Theory Alliance mentoring workshop on November 20. Fully free & virtual!
Theme: Harnessing AI for Research, Learning, and Communicating
Participants will have advanced ChatGPT capabilities (TY @OpenAI!)
Ft @Aaroth @risteski_a @RICEric22 & more
Congratulations to my PhD student Hirad Daneshvar for successfully defending his thesis on "Trustworthy Graph Neural Networks for Healthcare Predictive Modelling." Bravo Dr. @Hirad_Daneshvar ! @FEAStorontometu @TMU_ecbe @VectorInst
Check our recent paper on a precise and efficient method of quantifying GNNs predictive uncertainty, appearing at AIiH'25: link.springer.com/chapter/10….
@Hirad_Daneshvar @TMU_ecbe @VectorInst @FEAStorontometu
Conformal Prediction (CP) is a nice and (almost) assumption-free framework for quantifying uncertainty by constructing prediction sets with finite-sample coverage guarantees. @thehdmx
However, the set generated by the classical CP, with a rigid score function, is unable to capture the varying importance of the prediction set, which has adverse implications for confidence measurement in image segmentation or risk control in navigating robots. @CancerCenterAms
We addressed this in our paper openreview.net/pdf?id=uQaPr1… in
@midl_conference by devising a spatially aware CP function that incorporates the varying importance of regions in a segmentation task with applicability to other fields. #MIDL2025 @VectorInst @researchTMU @TMU_ecbe
Another well-organized and successful ML Security and Privacy workshop. Active TAILab participation with posters on Uncertainty quantification (@thehdmx @Hirad_Daneshvar) and Certified robustness (@MrezaMaleki_95) . Thanx to @VectorInst @thegautamkamath @TMU_ecbe
thanks @dev1cenet . here are our 2 recent papers on this topic:
arxiv.org/pdf/2506.20046
openreview.net/pdf?id=uQaPr1…
and two more from last year:
arxiv.org/pdf/2406.10787
arxiv.org/pdf/2306.00876
Reza Samavi retweeted
How can a robot traverse crowds of people? Checkout this press release about our crowd navigating robot.
tum.de/en/news-and-events/al…
Reza Samavi retweeted
Jack, a wheeled #robot from our university, navigates crowds by predicting people’s paths and adjusting 10 times per second. Rather than stopping, it adapts smoothly to avoid collisions: go.tum.de/528400
#RoboticsInstituteGermany #AutonomousVehicles #AI
📷A. Eckert
Reza Samavi retweeted
17 platforms where you can begin cybersecurity🚀
1. HackXpert - Free labs and training.
2. TryHackMe - Hands-on exercises and labs.
3. CyberSecLabs - High quality training labs.
4. Cybrary - Videos, labs, and practice exams.
5. LetsDefend - Blue team training platform.
6. Root Me - Over 400 cybersecurity challenges.
7. RangeForce - Interactive and hands-on platform.
8. Certified Secure - Loads of different challenges.
9. Vuln Machines - Real world scenarios to practice.
10. Try2Hack - Play a game based on the real attacks.
11. TCM Security - Entry level courses for cybersecurity.
12. EchoCTF - Train your offensive and defensive skills.
13. Hack The Box - Cybersecurity training platform.
14. Vuln Hub - Material for hands-on experience.
15. OverTheWire - Security concepts via challenges.
16. PentesterLab - Learn web-app penetration testing.
17. PortSwigger Web Security - General learning.
credit-Daniel Kelley
Reza Samavi retweeted
i like both AI and end-to-end encryption. can i have them all? we examine this question carefully in <How To Think About End-To-End Encryption and AI: Training, Processing, Disclosure, and Consent> by Knodel, Fabrega, Ferrari, Leiken, Hou, Yen, de Alfaro, Cho and Sunoo Park.
Reza Samavi retweeted
Happy Robotic Holidays from the Learning Systems and Robotics Lab! Episode 1: Delivering Gifts.
Video: youtu.be/hqZ4NJLLot0
ALT Video: https://youtu.be/hqZ4NJLLot0
Reza Samavi retweeted
Checkout our robot getting ready for the holidays using SICNav to move around!
Video: youtu.be/hqZ4NJLLot0
Paper: sepehr.fyi/projects/sicnav
Happy Robotic Holidays from the Learning Systems and Robotics Lab! Episode 1: Delivering Gifts.
Video: youtu.be/hqZ4NJLLot0
ALT Video: https://youtu.be/hqZ4NJLLot0
Congratulations Dr. Mini Thomas! A long successful journey and a big celebration day! @McMasterEng @McMasterU @TMU_ecbe
Reza Samavi retweeted
Our latest work Safe and Interactive Crowd Navigation (SICNav) has been accepted to @ieeeTRO! SICNav jointly solves planning and prediction while satisfying safety consts! Thanks @florian_shkurti @angelaschoellig
preprint & code: sepehr.fyi/projects/sicnav
youtu.be/BY23UXwM1sM
Reza Samavi retweeted
.@thegautamkamath and his team have shown that the best robust estimators often translate to the best private estimators. This connection has important implications for designing statistical methods. Check it out below⬇️
vectorinstitute.ai/robustnes…
