Simon Guiroy

Our paper (accepted at @CoLLAs_Conf) is out! arxiv.org/abs/2208.02377 We show that Meta-Learning generalization to novel OOD task distributions can be inferred from the neural activation dynamics from a few unlabeled examples, and we propose Activation-Based Early-Stopping (ABE).

Tu veux agir contre la crise climatique? Une grande campagne se met en branle à travers le Québec pour s’attaquer au premier problème climatique au pays : Les énergies fossiles. Va sur urgence-climatique.ca pour faire pression sur ton élu.e fédéral.e en quelques clics.

Le Canada émet la majorité de ses GES par ses exportations de pétrole et de gaz vers l’étranger. Mais le gouvernement refuse de les prendre en compte. Cette année : exigez, en trois clics, que ces émissions soient incluses dans le Plafonnement des GES👉front-etudiant.ca/agis

Authors : @GuiroySimon ,@chrisjpal , Gonçalo Mordido , @apsarathchandar . Institutions : @Mila_Quebec , @UMontreal , @polymtl

Our paper (accepted at @CoLLAs_Conf) is out! arxiv.org/abs/2208.02377 We show that Meta-Learning generalization to novel OOD task distributions can be inferred from the neural activation dynamics from a few unlabeled examples, and we propose Activation-Based Early-Stopping (ABE).

Scaling Laws for the Few-Shot Adaptation of Pre-trained Image Classifiers abs: arxiv.org/abs/2110.06990 scaling the training set size improves few-shot performance and can be well approximated via power laws

Hello World! We are finally on Twitter! @ChandarLab is a research group based in @polymtl and @Mila_Quebec led by @apsarathchandar. Our work focuses on #LifelongLearning, #DeepLearning, #ReinforcementLearning, and #NLProc. Follow us for the latest news on our work!

We are grateful to the support from @polymtl, @MILAMontreal, @CIFAR_News, and @NSERC_CRSNG!

URGENT. Make noise about Belarus. Tell your friends in the media. Rubber bullets and flash grenades are used against people in the streets.

I am teaching a new graduate-level Machine Learning course @polymtl / @MILAMontreal in Fall 2020 - sarathchandar.in/teaching/ml/fa…. The course is online. If you are a student in any of the Quebec Universities, you can register for the course here: bci-qc.ca/en/students/ae…

@david_rolnick @mcgillu @MILAMontreal Glad to have you with us David! :)

Excited to announce I am starting this fall as Assistant Professor of Computer Science at @mcgillu /@MILAMontreal. I'll be taking grad students and postdocs both in deep learning theory and in machine learning applied to climate change - please get in touch if interested!

@ericneumannpsyc It's actually freaking awesome

Lots of content about how hard, unrewarding, painful, exhausting a PhD is. Good to know, but also a bit terrifying for people like me who start their PhD... in a month! So, what's something about pursuing a PhD that turned out BETTER than expected?

Climate activists encampment in the heart of Montreal globalnews.ca/video/7229222/…

BREAKING: We're receiving reports that at least one protester was abducted off the streets today by unmarked “officers” — this time in New York City. These dangerous, abusive, and indefensible actions must stop. Law enforcement must be held accountable.

The new @MILAMontreal research blog is out!

Registrations are now open for our free virtual event: eventbrite.com/e/mais-2020-ti… Register to have access to the interactive poster session and webinar on September 11th!

I'm increasingly concerned about the exponential cost of #AI. #gpt3 from @openAI is amazing, but training it takes 355 years on a single GPU and costs $4.6M on current GPU cloud clusters. Do we really want a world where only a few rich companies can use SoA #AI? #DeepLearning

Come tomorrow at #ICML 4th Lifelong Learning Workshop! I'll be online all day from 7AM EDT, to present and discuss my work: "Towards an Unsupervised Method for Model Selection in Few-Shot Learning" See 5min video: youtu.be/RAOhJTB4Clw Looking forward to those discussions :)

@MILAMontreal @chrisjpal @vikasverma1077

Our paper is at #ICML2020 - Lifelong Learning Workshop! Do we need labeled data from held-out classes for model selection in Few-shot Meta-Learning? Our method doesn’t. Using all available data for training, results show it can outperform meta-validation. youtu.be/RAOhJTB4Clw

Excited to release Torchmeta! A collection of extensions and data-loaders for few-shot learning & meta-learning in @PyTorch github.com/tristandeleu/p… / pip install torchmeta Paper: arxiv.org/abs/1909.06576