Patrice Bechard

🧠 What can the brain tell us about AI and vice versa? Guillaume Lajoie, a Canada CIFAR AI Chair at Mila - Quebec Artificial Intelligence Institute and professor at the Université de Montréal, works at the intersection of these two fields to uncover the fundamental principles of intelligence and build more intelligent AI systems. Watch as he explains the impact of his work, from better learning mechanisms in AI systems to the treatment of neurological diseases. #AI #CanadianAI #NeuroAI

55

1 Comment

𝐀 𝐫𝐞𝐜𝐞𝐧𝐭 𝐈𝐁𝐌 𝐚𝐧𝐝 𝐌𝐨𝐫𝐧𝐢𝐧𝐠 𝐂𝐨𝐧𝐬𝐮𝐥𝐭 𝐬𝐮𝐫𝐯𝐞𝐲 𝐟𝐨𝐮𝐧𝐝 𝐭𝐡𝐚𝐭 𝟗𝟗% 𝐨𝐟 𝐞𝐧𝐭𝐞𝐫𝐩𝐫𝐢𝐬𝐞 𝐝𝐞𝐯𝐞𝐥𝐨𝐩𝐞𝐫𝐬 𝐚𝐫𝐞 𝐧𝐨𝐰 𝐞𝐱𝐩𝐥𝐨𝐫𝐢𝐧𝐠 𝐨𝐫 𝐝𝐞𝐯𝐞𝐥𝐨𝐩𝐢𝐧𝐠 𝐀𝐈 𝐚𝐠𝐞𝐧𝐭𝐬. 𝟐𝟎𝟐𝟓 𝐡𝐚𝐬 𝐚𝐥𝐫𝐞𝐚𝐝𝐲 𝐛𝐞𝐞𝐧 𝐜𝐚𝐥𝐥𝐞𝐝 “𝐭𝐡𝐞 𝐲𝐞𝐚𝐫 𝐨𝐟 𝐭𝐡𝐞 𝐀𝐈 𝐚𝐠𝐞𝐧𝐭” 𝐚𝐧𝐝 𝐰𝐞 𝐚𝐫𝐞 𝐞𝐱𝐩𝐞𝐫𝐢𝐞𝐧𝐜𝐢𝐧𝐠 𝐚 𝐬𝐡𝐢𝐟𝐭 𝐟𝐫𝐨𝐦 𝐬𝐭𝐚𝐧𝐝𝐚𝐥𝐨𝐧𝐞 𝐦𝐨𝐝𝐞𝐥𝐬 𝐭𝐨 𝐚𝐠𝐞𝐧𝐭𝐢𝐜 𝐬𝐲𝐬𝐭𝐞𝐦𝐬. To understand what this means for software engineers and architects, for Deep Engineering #6, I spoke with @Imran Ahmad, data scientist at Canada’s Advanced Analytics Solution Center, visiting professor at Carleton University, and author of 50 𝑨𝒍𝒈𝒐𝒓𝒊𝒕𝒉𝒎𝒔 𝑬𝒗𝒆𝒓𝒚 𝑷𝒓𝒐𝒈𝒓𝒂𝒎𝒎𝒆𝒓 𝑺𝒉𝒐𝒖𝒍𝒅 𝑲𝒏𝒐𝒘 (@Packt, 2023). 𝐖𝐞 𝐜𝐨𝐯𝐞𝐫𝐞𝐝: 🧠 Why orchestration—not just output—defines real intelligence ⚙️ How classical algorithms support modern AI workflows 📏 Why optimization is essential  🔁 How algorithm choice shapes architecture in scalable systems Imran brings a practitioner’s view shaped by public-sector constraints, cloud-scale deployments, and deep algorithmic thinking. His next book, 30 𝑨𝒈𝒆𝒏𝒕𝒔 𝑬𝒗𝒆𝒓𝒚 𝑨𝑰 𝑬𝒏𝒈𝒊𝒏𝒆𝒆𝒓 𝑺𝒉𝒐𝒖𝒍𝒅 𝑲𝒏𝒐𝒘, 𝒊s due later this year. Rea𝐝 𝐭𝐡𝐞 𝐟𝐮𝐥𝐥 𝐢𝐬𝐬𝐮𝐞 𝐚𝐧𝐝 𝐢𝐧𝐭𝐞𝐫𝐯𝐢𝐞𝐰: 𝐡𝐭tps://https://lnkd.in/gYs9nttf #AIagents #Algorithms #SystemDesign #Scalability #DeepEngineering

16

We were thrilled to see a packed room of engaged AI practitioners at Amanda Milberg’s session at Toronto Machine Learning Summit (TMLS) yesterday! Amanda covered essential considerations for productionizing GenAI in regulated industries, including: ☑️ The building blocks of AI agents ☑️ What an MCP server is ☑️ A step-by-step guide to self-hosting AI agents The response was incredible - thank you to everyone who joined. 💡 If you missed it and would like access to the recording, just reach out.

16

1 Comment

🚨 Quantum Experiment Success from Canada! 🇨🇦🚨 THE UNRAH EFFECT! 🔗 Want to experiment on Qiskit? Here is the link to Github: https://lnkd.in/egPUH4A5 . We're thrilled to announce that Dr. Zuhair Ahmed, affiliated with the Centre of Excellence for Technology, Quantum, and AI Canada (CETQAC), has successfully simulated the Unruh Effect on IBM’s real quantum hardware — a first-of-its-kind achievement globally. The Unruh Effect, a quantum-relativistic phenomenon predicted in 1976, has long remained in the realm of theory. Dr. Ahmed’s innovative quantum circuit—leveraging Qiskit Runtime and IBM’s superconducting qubits—transforms this abstract concept into experimental reality, advancing our understanding of spacetime, entropy, and quantum thermodynamics. 📊 Key Results: Purity: 0.5 ✔️ Entropy: 1.0 ✔️ Entanglement Entropy: 1.0 ✔️ Fidelity vs. Vacuum: 0.46 ✔️ This simulation proves the potential of quantum computing to explore the intersection of relativity and quantum mechanics, and places Canada at the forefront of quantum research. 💡 “By simulating the Unruh effect, we’re opening new doors to the universe’s deepest secrets,” says Dr. Ahmed. Read the Pre-print on ResearchGate: https://lnkd.in/encszjNY #QuantumComputing #UnruhEffect #CETQAC #QuantumBreakthrough #QuantumRelativity #IBMQuantum #MadeInCanada #ScientificMilestone #Physics #Innovation #ResearchExcellence #Qiskit #QuantumThermodynamics #QuantumGravity #CanadaQuantumLeadership

2

We’re excited to launch the Microsoft-Northwestern-WITNESS deepfake detection benchmark — a collaborative, open-source dataset of over 50,000 AI-generated and real-world manipulated media artifacts. 🧠 Expert-annotated 🌍 Real-world, high-stakes examples 🧪 Includes adversarial attacks 🚫 Evaluation-only (not for training or commercial use) Our goal: to set a high bar for detection tools and support developers, journalists, and human rights defenders worldwide. We invite the AI community to help us keep this benchmark up to date. Collaboration between Microsoft AI for Good Lab: Thomas Roca, Juan M. Lavista Ferres, Pengce Wang Kevin White Northwestern Security & AI Lab: Chongyang GAO, Isabel Gortner, Marco Postiglione, Sarit Kraus, and me WITNESS: Mahsa Alimardani, shirin anlen, Zuzanna W. and Sam Gregory To find the dataset and associated paper, visit: 🔗 https://lnkd.in/gP-_E6Aw #AI #Deepfake #ResponsibleAI #OpenSource #HumanRights #Cybersecurity #NationalSecurity #InternationalSecurity #Defense #Defence #GlobalPublicGood #AIForGood #MachineLearning #GenerativeModel

138

9 Comments

Unlock the power of AI with the Neural Networks course from the University of Waterloo. Dive into RNNs, CNNs, transformers, and hands-on Python training in Jupyter Notebooks—designed to fit your schedule with flexible pacing and expert support. The course starts September 18. Join to build your technical portfolio, whether you're aspiring to become a data analyst or simply want to stay AI-savvy.  https://lnkd.in/gP7VB3Ch #MachineLearning #NeuralNetworks #AI #WatSPEED #UniversityOfWaterloo

5

“ IQC has also spun out more than 20 quantum startups and 40 per cent of its researchers are involved in commercialization.” Researchers at the Institute for Quantum Computing (IQC) at the University of Waterloo are doing more than tackling trailblazing theoretical and applied researc… Source: Waterloo News https://buff.ly/CPCUxKB

7

I have recently finished reading TRM paper [Samsung AI Lab (SAIL) Montréal] Tiny Recursive Models on ARC-AGI-1: Inductive Biases, Identity Conditioning, and Test-Time Compute and here's what i found: Large transformer models often push 100M–1B+ parameters just to gain incremental reasoning improvements, yet still plateau around ~40% accuracy on structured tasks like Sudoku, maze planning, and ARC-AGI-style problems. TRM, with only ~7M parameters, reaches ~80% accuracy on similar benchmarks by enforcing recursive, step-wise reasoning — effectively delivering 2× the accuracy with a fraction of the scale. That gap is a HUGE DEAL: it shows that structured recursion can outperform brute-force parameter growth on logic-heavy tasks. Tiny Recursive Models (TRM) changed how I think about reasoning architectures. After reading the paper, here’s what stood out to me: 1. Recursion instead of scale TRM shows that you don’t always need massive parameter count to solve reasoning heavy tasks. Instead, you reuse a small network recursively. The same module solves progressively refined subproblems. That’s compute reuse, not brute force scale. 2. Depth via time, not parameters Instead of stacking 100+ layers once, TRM applies a small model multiple times. It trades width for iterative refinement. Think: controlled reasoning steps rather than one-pass prediction. 3. Why this matters for HRM HRM (Hierarchical Reasoning Model) separates: High-level planning (H-module) Low-level execution (L-module) The improvement TRM introduces is structural discipline: Iterative refinement makes the planning phase more stable. Recursive reuse reduces overfitting from large monolithic planners. Deep supervision across steps improves intermediate reasoning signals. In short: TRM strengthens HRM by making reasoning compositional instead of oversized. This feels like a shift from “bigger transformer” → “structured computation.” Scaling laws are not the only path forward. Structured recursion is another axis.

4

Pleased to share our last research entitled "Noise-Optimized Conditional Diffusion for Domain Adaptation (NOCDDA)", which has just been accepted in IJCAI, Montréal, August 2025 Domain adaptation is essential to enable models to generalize across shifting real-world conditions—without the burden of labeling every new domain. In this work, we tackle a core challenge in unsupervised domain adaptation (UDA) — the limited availability of high-confidence pseudo-labeled target samples. We introduce a conditional diffusion framework that bridges generative modeling and cross-domain decision-making, with a strong emphasis on class-aware noise optimization. The preprint can be found here: https://lnkd.in/eiCXRED6 The International Joint Conference on Artificial Intelligence (IJCAI) is one of the premier global conferences in AI, bringing together top researchers from academia and industry since 1969. With a highly competitive review process and a strong focus on cutting-edge innovations across all areas of artificial intelligence, IJCAI has remained the premier conference bringing together the international AI community. This year’s edition will be held in Montréal, Canada, continuing its tradition of fostering high-impact AI research and collaboration.

28

Many adaptation methods for large language models still rely on heavy compute and don’t fully solve issues like hallucination — which is especially problematic in technical domains such as engineering. In our latest work at McGill University under the supervision of Professor Fiona Zhao, we introduce the Small Language Graph (SLG): a lightweight, modular system that connects small, fine-tuned language models into a graph of domain-specific experts. Key findings compared to a stand-alone, fine-tuned larger language model: 🔹 3× improvement in Exact Match — a metric that compares generated text to ground truth, measuring exact correspondence 🔹 1.7× faster fine-tuning time 🔹 Lower resource requirements for both training and inference What’s exciting is the potential for distributed AI. Inspired by systems like torrents, the expert nodes represented by small language models could eventually run on everyday devices — acting as nodes in a network. This could reduce the need for massive centralized compute clusters and make generative AI more accessible and scalable. Acknowledging Meta and Yann LeCun for their support of open-source AI, which enabled this research. #AI #LLM #NLP #LanguageModels #DistributedAI #DomainAdaptation #EngineeringAI #EdgeAI #SLG https://lnkd.in/dcA9VZAQ

21

Innovasea partners with Mila to advance AI innovation Innovasea announced today a new partnership with Mila, the Quebec Artificial Intelligence Institute. Together, the two organisations aim to advance AI development for sustainable fish tracking, protection and production. https://lnkd.in/dWbjwv_i

3

Quantum Computing (QC) has tremendous potential - Starting from first principles creating a computer that follow quantum mechanics make so much sense. Wonderful research will emerge from the combination of AI and QC in the coming years - I highly recommend applying (Consider making a transition to the field if you are not already from QC background)!

4

Canada’s quantum future starts now! FABrIC’s Quantum Program is helping Canadian industry prepare for the arrival of commercially viable quantum technologies. From hands-on access to quantum computers through the Quantum Computing Sandbox to expert-led training on hardware, software, and applications—this is your chance to lead in the next wave of innovation. 🔗 Learn more and get ready: https://loom.ly/8AGyD8I #QuantumReadiness #InnovationCanada #FutureTech Institute for Quantum Computing Quantum Industry Canada • Industrie Quantique Canada Photons Canada/ CPIC Quantum City National Research Council Canada / Conseil national de recherches Canada Quantum Algorithms Institute UBC Blusson Quantum Matter Institute (Blusson QMI) Quantum BC Pacific Economic Development Canada | Développement économique Canada pour le Pacifique DIGITAL Invest Ontario Invest Ottawa Innovate BC Alberta Innovates Invest Nova Scotia Distriq, Zone Innovation Quantique de Sherbrooke Ministère de l’Économie, de l’Innovation et de l’Énergie du Québec (MEIE) Québec Quantique par Distriq

19

1 Comment

🎉 RAML Lab Congrats! 🎉 We’re excited to announce that David Martin’s paper, “Reducing Representation Bias through Fairness-Driven Sampling in Contrastive Learning,” has been accepted at the Canadian AI Conference 2026 (Round 1)! This work tackles a key challenge in contrastive learning: how standard random sampling can amplify representation bias, causing overrepresented groups to dominate training while underrepresented groups receive limited exposure. The paper introduces a fairness-driven sampling strategy that uses latent similarity structure to guide contrastive pair selection without relying on demographic labels. The proposed method improves representation of underrepresented latent subgroups while maintaining competitive model accuracy, with promising implications for fairer downstream applications such as facial recognition, clinical diagnostics, and language models. Excellent work and a strong contribution to fair and responsible self-supervised learning. Congratulations to David Martin and the entire RAML Lab team! Special thanks to our director, Dr. Blessing Ogbuokiri, for the guidance and leadership.👏 #RAMLLab #CanadianAI2026 #AIFairness #AISafety #ResponsibleAI #ContrastiveLearning #SelfSupervisedLearning #MachineLearning #BiasMitigation #TrustworthyAI

16

1 Comment

Does Reinforcement Learning really make LLMs “reason better”?This paper finds RL often makes models more consistent, but not necessarily more capable of new reasoning. It’s great at 1-shot success, but base models catch up when sampling more. Sometimes we’re just optimizing within what the model already knows.

7

Check out our ICML paper “Demystifying Singular Defects in Large Language Models”. We find a way to tell whether a model is fine-tuned from another! https://lnkd.in/ebsmmSAv #ICML #LLM #Vancouver

61

Anthropic just released Claude 4, including: - Opus 4: their most powerful model yet - Sonnet 4: the faster, cheaper alternative for everyday tasks While the capabilities are promising, here are some thoughts. 🔹 Pricing is a major friction point — $75 per million output tokens for Opus is steep compared to: • GPT-4o at $15 per million tokens • Gemini 2.5 pro at $10 per million tokens (flash is even cheaper) • Mistral models: many are free and open-weight 🔹 Sonnet 4 over 3.5 — is it truly next-gen or just iterative? Eager to see how Claude 4 performs in complex reasoning, long-context tasks, and clean code generation. Read good comments from some of the early adopters. In an ecosystem increasingly split between closed and open models, transparency, access, and cost will shape what (and who) wins. https://lnkd.in/gTquwwEk

18

1 Comment

Advancing Research Workflows with Agentic AI Systems: I am excited to share my recent project called Agentic ArXiv Research Report Generator, an AI system designed to autonomously analyze academic papers from ArXiv and produce structured, accurate, and well written research reports based on user provided documents. The core idea behind this project is to model the workflow of a real research team using a set of collaborating agents that each perform distinct roles while working together in a coordinated reasoning loop. The system includes a planning agent that interprets user intent, identifies the main contributions of the paper, and outlines the structure of the report. A writer agent then generates the analytical narrative using the retrieved context. A reviewer agent evaluates clarity, factual correctness, flow, and overall cohesion and sends feedback that improves the final report, allowing the system to refine its output with minimal human intervention. Technically, this project demonstrates work in agentic AI systems, retrieval augmented generation workflows, and autonomous task orchestration using LangGraph. It incorporates semantic indexing and contextual retrieval through LlamaIndex, and performs local inference using LLMs integrated via Ollama to enable privacy-preserving and cost-efficient model execution. The project involved careful design of modular Python components, state management logic for coordinated agent behaviors, and prompt strategies that guide multi step reasoning across different stages of the workflow. It strengthened my skills in designing intelligent pipelines where planning, content generation, and quality assurance operate as interconnected processes rather than isolated tasks. This project reflects my growing focus on AI engineering for research automation and knowledge workflows. I am excited about the potential of agentic AI to support students, researchers, and analysts by reducing time spent summarizing complex papers and increasing time spent on interpretation, creativity, and higher level thinking. I look forward to exploring collaborations and further development in areas such as multi agent systems, context aware reasoning, and scalable research support tools. 🔗 You can explore the code and sample generated report here: https://lnkd.in/d8ky5A4y #AIEngineering #AgenticAI #LangGraph #LlamaIndex #Ollama #RAG #MultiAgentSystems #ResearchAutomation #ArtificialIntelligence #MachineLearning

12

Nearly 300 academic papers accepted at top AI conferences contain citations that don't exist. 🫢 Researchers at NAIST just published an analysis of 17,000+ papers from ACL, NAACL, and EMNLP conferences. What they found is alarming: "HalluCitations", completely fabricated references to papers that were never written, are exploding. The trend: - 2024: 20 papers with fake citations - 2025: 275 papers - EMNLP 2025 alone: 154 papers, with over 100 in main conference tracks While this represents less than 2% of papers overall, the 13x year-over-year increase signals a trend that demands attention before it becomes systemic. And these aren't obvious fakes. They're buried among 30-50 real citations. They have plausible-looking arXiv IDs. Plausible authors. Plausible titles. Even reviewers with domain expertise can't catch them all. But here's the nuance they emphasize: HalluCitations don't always mean someone used ChatGPT to write their paper. Sometimes authors copy citations from secondary sources without checking. Sometimes it's human error. The solution isn't punishing authors. It's building better pre-submission checks. The researchers are calling for HalluCitation detection tools to be integrated into standard toolkits like ACL pubcheck. What does this mean for us building AI systems? → Verification layers matter. If your system generates citations, you need citation verification. And this is crucial not only for citation, but also for any critical information that your system generates. 🔗 Paper: https://lnkd.in/egVFeJuk Have you ever caught a hallucinated citation in a paper you were reviewing or reading? #AI #MachineLearning #AcademicIntegrity #LLM #ResponsibleAI #Research

21

1 Comment