Dominique Beaini

🌟 Reconnecting with a long-standing Friend and Discussing the Future of AI in Computer Vision & Medical Image Analysis 🔍I recently had the incredible opportunity to host Professor Tal Arbel and her students from McGill University and Mila - Quebec Artificial Intelligence Institute at TheraPanacea. Tal is not only an outstanding colleague and a great asset to the community but also a fantastic friend. Her energy, motivation, and engagement extend far beyond the scientific realm.  🩺 We discussed how AI and machine learning have revolutionized medical image analysis, enhancing diagnostic accuracy and predictive analytics. Technologies like deep learning are transforming healthcare, leading to better patient outcomes and more efficient workflows in fields such as oncology and neurology. 🚀 Our primary focus, however, was on future directions. We explored concepts like personalized medicine, which aims to tailor treatments based on individual imaging data, moving towards holistic treatment pathways. We also emphasized the importance of collaborative efforts in an academic-industry startup setting, involving clinical experts, advanced algorithms, and data management to drive innovation. The synergy between academia and industry will be pivotal in shaping the future of healthcare. #AI #MedicalImaging #HealthcareInnovation #FutureOfMedicine #AcademicConnections

111

5 Comments

Exciting New Collaboration Published! 🎓 I'm thrilled to announce the release of our new paper titled "Optimizing Encoder-Only Transformers for Session-Based Recommendation Systems", a joint effort between Air Liquide and Polytechnique Montreal. This work is the foundation of the Master’s thesis by first author Anis Redjdal, M.Sc. In the paper, we introduce Sequential Masked Modeling, a novel approach to improve the performance of encoder-only transformers in session-based recommendation systems, alongside architectural improvements inspired by modern LLMs. By enhancing how transformers capture sequential dependencies, our method achieves state-of-the-art results in next-item prediction across several benchmark datasets. 🔗 https://lnkd.in/eeh6BvUg

46

3 Comments

✨ Hello LinkedIn Community! Have you contributed to open-source software (OSS) projects as a designer or a developer during the past year? We are inviting you to take part in an interview study conducted by researchers from Polytechnique Montréal and McGill University. 🎯 Study Goal: This study aims to improve OSS processes and tools by exploring new ways to involve designers in OSS communities through innovative design approaches. ⏳ Time Commitment: Approximately one hour per session, with two sessions in total. 🌀 Process: You will participate in two individual interview sessions, where we will explore your experiences contributing to OSS projects and ask for your reflections based on fictional worlds, we created to inspire discussion on OSS design and usability. The interviews will be conducted virtually (via Teams or Zoom)  and will be video recorded for accuracy. 🎖️ Compensation: You will receive a total of $60 CAD for your participation in both sessions. 🔏 Confidentiality: Your privacy is a priority. Your information and identity will remain confidential and accessible only to the research team. ✉️ Contact: Express your interest or ask questions by reaching out to Rozhan Hozhabri Nezhad ([email protected]) or Jinghui Cheng ([email protected]). Please share this opportunity with your peers and friends who are OSS designers or OSS developers. Your contribution and network will be invaluable in making this study a success!

24

1 Comment

Are you thinking about participating in the Quantum Resource Estimation challenge, but don't know where to start? Check out the video tutorial on using Azure Quantum Resource Estimator! https://lnkd.in/gqGUhzp5

50

1 Comment

There's only 5 days left to submit your paper! Remember, we only a short 1-page abstract

2

In our latest paper with Jean-Baptiste Waring, B.Eng., we generated a dataset on `ibm_quebec` the 127-qubit IBM Quantum System One located in Canada. The dataset contains more than 4000 experiments of long-range entanglement path qubits from 0 to 100+ qubits. The dataset and the code can be found here: https://lnkd.in/eussBW9G Paper: https://lnkd.in/ekb9Vqja A special congratulations to Jean-Baptiste Waring, B.Eng. for this repo and the quality of the experiments. #quantumcomputing

47

1 Comment

Meta AI Unveils EvalGIM: A Cutting-Edge Library for Evaluating Generative Image Models Researchers from Meta’s FAIR lab, alongside collaborators from Mila Quebec AI Institute, Université Grenoble Alpes, Inria CNRS, McGill University, and Canada CIFAR AI Chair, have introduced EvalGIM—an advanced machine learning library designed to standardize and enhance the evaluation of text-to-image generative models. EvalGIM bridges critical gaps in generative model assessment by supporting diverse metrics, datasets, and visualization tools, enabling robust and adaptable evaluations. Its standout feature, Evaluation Exercises, synthesizes insights into key performance trade-offs, such as quality versus diversity and demographic representation. With a modular architecture, EvalGIM ensures seamless integration of new evaluation components, staying relevant in the rapidly evolving AI landscape. The library incorporates widely-used datasets, including MS-COCO and GeoDE, to evaluate model performance across various geographic regions. It also features prompt-only datasets like PartiPrompts and T2I-Compbench, enabling diverse text input testing. Compatible with tools such as HuggingFace diffusers, EvalGIM supports benchmarking from early training stages to advanced iterations. Notable features include distributed evaluations for accelerated analysis across compute resources and hyperparameter sweeps for in-depth exploration of model behaviors. Researchers can also integrate custom datasets and metrics effortlessly, making EvalGIM a versatile solution for generative image model evaluation. For more details: Research Paper :  https://lnkd.in/gvvmwSiJ GitHub Repository : https://lnkd.in/gRBYzW-z #MetaAI #EvalGIM #MachineLearning #GenerativeAI #TextToImage #AIResearch #ComputerVision #ModelEvaluation #HuggingFace #MSCOCO #GenerativeModels #AIEthics #DataScience #ArtificialIntelligence #AIInnovation #OpenSourceAI #DeepLearning Umar Iftikhar

2

One of my funniest god old challenges to LLMs, now on GPT-4 ... hum... not yet. Me: I have three glasses stacked on top of each other and want to swap the bottom one with the top one. What should I do ? ChatGPT To swap the bottom glass with the top one when you have three glasses stacked on top of each other, you can follow these steps: Remove the top glass: Carefully lift the top glass off the stack and set it aside. Remove the middle glass: Lift the middle glass off the stack and set it aside. Move the bottom glass to the top: Lift the bottom glass and place it where the top glass was. Re-stack the middle glass: Place the middle glass on top of the new bottom glass. Re-stack the original top glass: Finally, place the original top glass on top of the stack. Now, you have successfully swapped the bottom glass with the top one.

126

30 Comments

Today, a small tutorial for my students in API6322, GSPIA, University of Ottawa - How to use generative artificial intelligence to 'prompt' a large language model to do a scatter plot using pandasai, then how to do it more traditionaly using python matplotlib library and then using Excel. #datavisualization #python #excel #generativeai

23

Last week, Alexandre Dubray presented "Anytime Weighted Model Counting with Approximation Guarantees for Probabilistic Inference", joined work with Pierre Schaus, at the Constraint Programming conference. In this work, we show how to modify a model counting solver to calculate an approximate model count, with approximation guarantees at any moment the solver is stopped. We believe such a solver can be useful for inference tasks on many probabilistic models: Bayesian networks, neurosymbolic models, and more. https://lnkd.in/eg8FQcBc

21

Excited to share our latest publication in Computers in Biology and Medicine. In collaboration with the BC Children's Hospital Research Institute, this study introduces BRI-Net, an interpretable deep-learning model that leverages adversarial training to learn an unbiased representation of data, helping identify genetic markers associated with specific phenotypes. Using genomic data from a cohort of 362 pediatric cancer patients treated with cisplatin, recruited by the Canadian Pharmacogenomics Network for Drug Safety, our model identified two statistically significant SNPs associated with cisplatin-induced ototoxicity: CERS6 rs13022792 and TLR4 rs10759932. To explore our methodology and findings in detail, check out the full paper at https://lnkd.in/gEn4fZb9. A huge thanks to my amazing co-authors and collaborators, especially Bahareh Kashani, for making this possible!

66

4 Comments

This is a great writeup by Hugging Face on how to improve the output performance of LLMs by spending more compute resources when doing inference (what is being called test-time compute), as opposed to the classic neural net paradigm where you spend a huge amount of compute in training and then a fixed amount of compute on each subsequent inference. The analogy here is to how humans "think harder" for hard problems, dwelling on partial answers and iterating before producing a final answer. The motivation is to move further from pure next-toke prediction, for which the analogy is answering off the top of your head. I think this is a nice alternative to agentic AI systems, which involve setting up a series of models with different roles that review and respond to each others' outputs. Agentic AI also represents a way of iteratively spending more compute on generating answers, when required. Next-token prediction fundamentally cannot provide backtracking, even with prompts such as "think step by step", and so it's easy for them to get stuck with a bad output if they head off in the wrong direction initially. Tree search and agentic AI (if it includes iteration) are beginning to introduce backtracking strategies into LLM outputs.

16

1 Comment

🚀 Excited to present our technical paper at IEEE Quantum Week 2024 in Montréal! 🚀 Join me on Wednesday, for my presentation on "A Hybrid Quantum-Classical Physics-Informed Neural Network Architecture for Solving Quantum Optimal Control Problems" in "Quantum Machine Learning and Neural Network Architectures" session. Our work demonstrates how innovative neural network architectures can bridge the gap between quantum and classical systems to tackle complex control challenges. Looking forward to insightful discussions! See you there! 👋 #IEEEQuantumWeek2024 #QuantumMachineLearning #NeuralNetworks #QuantumComputing

7

2 Comments

NEW... How do AI chatbots work? : The Blankety-Blank unplugged activity. As I was thinking about teaching AI unplugged this week at the Playful Computing conference, here is a new unplugged activity to explore how chatbots (ie Large Language Models) generate text/answers, why they come up with different answers each time, why those answers may not be the best answers and why they may even by wrong. Ultimately they are just generating answers based on statistics. https://lnkd.in/eXPAyNY6 I will be adding it to our Learning about Machine Learning Pages. In short: have the class generate answers on post it notes to fill the blanks at the end of blankety-blank like statements such as "Birds _____", "Monkeys eat ______" , "Orangutans are _____" one statement at a time. Plot a bar chart by grouping the post it notes on the board for each statement in turn. Work out the most common answers from the bar chart. Choose between them randomly (toss a coin) to complete the sentence... Discuss how good, weak and wrong answers could be chosen. HEALTH WARNING: It has only been play tested in my head so far...

23

2 Comments

🚀 Exploring Multi-Modal Representation Learning Working in machine learning offers the incredible opportunity to apply similar methods to diverse datasets. The project in my internship at Mila - Quebec Artificial Intelligence Institute focuses on extracting valuable information from both images and point clouds, using data captured by UAVs. These high-resolution datasets provide an incredibly rich and detailed perspective, offering numerous opportunities for innovation. Despite the challenges posed by the size and multi-modality of the data, our goal is to enhance downstream tasks using self-supervised learning techniques. By leveraging these methods, we aim to create robust and scalable solutions for various applications. Here's a sneak peek of an orthomosaic image and a point cloud dataset from one of our locations. 🛰️ P.S. We have tens of thousands of such tiles, making this project both challenging and incredibly exciting! #MachineLearning #AI4Climate #UAV #Orthomosaic #PointClouds #MultiModal #RepresentationLearning

40

1 Comment

NiNo: A Novel Machine Learning Approach to Accelerate Neural Network Training through Neuron Interaction and Nowcasting Researchers from Samsung’s SAIT AI Lab, Concordia University, Université de Montréal, and Mila have introduced a novel approach known as Neuron Interaction and Nowcasting (NINO) networks. This method aims to significantly reduce training time by predicting the future state of network parameters. Rather than applying an optimization step at every iteration, as with traditional methods, NINO employs a learnable function to predict future parameter updates periodically. By integrating neural graphs—which capture the relationships and interactions between neurons within layers—NINO can make rare yet highly accurate predictions. This periodic approach reduces the computational load while maintaining accuracy, particularly in complex architectures like transformers. At the core of the NINO methodology lies its ability to leverage neuron connectivity through graph neural networks (GNNs). Traditional optimizers like Adam treat parameter updates independently without considering the interactions between neurons. NINO, however, uses neural graphs to model these interactions, making predictions about future network parameters in a way that reflects the network’s inherent structure. The researchers built on the Weight Nowcaster Networks (WNN) method but improved it by incorporating neuron interaction modeling. They conditioned NINO to predict parameter changes for the near and distant future. This adaptability allows NINO to be applied at different stages of training without requiring constant retraining, making it suitable for various neural architectures, including vision and language tasks. The model can efficiently learn how network parameters evolve by using supervised learning from training trajectories across multiple tasks, enabling faster convergence.... Read our article on this:https://lnkd.in/gSQFYtcG Paper: https://lnkd.in/g9rJ5JT2 GitHub Page: https://lnkd.in/gJPbT-kM

25

It was indeed a pleasure and an honor to be part of an amazing team in writing the international scientific report on AI safety, benefiting from the great expertise of contributors from various countries, under the chairmanship of Professor Yoshua. Congratulations to all the team members who participated in the writing of this report. Here is the link to the interim version of the report, which will be launched at the upcoming AI Safety Summit in Seoul: https://lnkd.in/dFET2BKX"

37

8 Comments