Maria Fernanda Rodriguez

We are happy to share PEAK, an AI assistant that helps optimize GPU kernels, recently published on arXiv by Angelica Moreira, Ph.D. and her collaborators. Angélica is a former PhD student from UFMG's Compilers Lab and is currently a Senior Research Software Engineer at Microsoft Research Redmond. PEAK is the result of a collaboration between Microsoft Research and Stanford University. The assistant optimizes GPU kernels using natural-language instructions: performance experts describe optimization steps (like how to tile loops or use shared memory) in simple language. These steps are then applied by an AI model and automatically checked for correctness and performance. PEAK supports three GPU backends: CUDA, HIP, and HLSL. It was evaluated on matrix multiplication kernels using 16 optimization steps. The results show that PEAK can match the best performance reported for the hardware, even in cases where no vendor-provided library is available. PEAK: A Performance Engineering AI Assistant for GPU Kernels Powered by Natural Language Transformations, by Muhammad Usman Tariq (Stanford University), Abhinav Jangda (Microsoft Research Redmond), Angélica Moreira (Microsoft Research Redmond), Madan Musuvathi (Microsoft Research), and Tyler Sorensen (Microsoft Research Redmond and University of California, Santa Cruz). Link: https://lnkd.in/dJCPqErw

108

6 Comments

Pattern recognition in hate speech detection: a neural network and ensemble approach TL;DR: New published paper! Hate speech on online platforms is a growing problem with significant social impacts. This work proposes an approach for binary classification of hate speech in Portuguese texts using machine learning and deep learning algorithms. The experiments were conducted on an annotated dataset, with textual representations generated by pre-trained GloVe word embeddings. The voting-based model, which combines the outputs of the base classifiers, achieved the best overall performance, reaching an F1-score of 0.76. The results demonstrate the effectiveness of neural networks, especially in capturing complex textual patterns, and highlight the potential of combined approaches for the hate speech classification task. This study reinforces the importance of exploring diverse architectures and preprocessing techniques tailored to the specific characteristics of the Portuguese language. https://lnkd.in/dqwwHH_Q

51

1 Comment

🇧🇷 I'm excited to share that our paper has been accepted to the main track of the World Conference on eXplainable Artificial Intelligence 2026 in Brazil! 🇧🇷 Great work with my supervisors Luca Longo and Miloš Savić! In this work, we explored our CORTEX (Cost-Sensitive Rule and Tree Extraction) algorithm on a real-world, highly imbalanced forest cover detection problem using Sentinel-2 cloud-free imagery. 🛰 This is the first application of CORTEX in two modeling paradigms (see diagram below), as inherently interpretable (ante-hoc) classifier and as a post-hoc surrogate model for MLP network. 📊 The results show that CORTEX produces smaller and simpler rules while maintaining competitive predictive performance compared to traditional decision tree model. 👀 For more details on CORTEX, check out our paper: https://lnkd.in/gzMj9vAw We're currently preparing an open-source python package, with GitHub repo to be released soon (happy to share it upon request in the meantime). If you're interested in contributing or collaborating, feel free to reach out! 👩🏻‍💻 #XAI #explainability #interpretable #MachineLearning #RemoteSensing #ForestDetection #Sentinel2 #research #AI #PhD

73

4 Comments

On a cold Saturday morning here in São Paulo, immersed in interpretability and explainability of Machine Learning models (XAI), I decided to create this repository. Born purely out of fun and educational purpose, it organizes scripts to train models (using Scikit-learn) and, more importantly, apply methods like LIME and SHAP to understand their decisions. Plus, unit testing. Although it started as a personal learning project, it can be effectively used to generate both local and global explanations for classification models all executable via the command line. :) https://lnkd.in/df2qub8A

8

📄We are happy to share this article from Vol. 6 : "End-to-End Neural Video Compression: A Review." ✍️ Authored by: Jiovana S. Gomes; Mateus Grellert; Fábio L. L. Ramos; Sergio Bampi Video compression is crucial in providing high-quality video services under limited network and storage capacities, traditionally achieved through hybrid codecs. However, as these frameworks reach a performance bottleneck with compression gains becoming harder to achieve with conventional methods, Deep Neural Networks (DNNs) offer a promising alternative. By leveraging DNNs’ nonlinear representation capacity, these networks can enhance compression efficiency and visual quality. Neural Video Coding (NVC) has recently received significant attention, with Neural Image Coding models surpassing traditional codecs in compression ratios. Therefore, this survey explores the state-of-the-art in NVC, examining recent works, frameworks, and the potential of this innovative approach to revolutionize video compression. 🔗 Continue reading on IEEE Xplore here👉 https://loom.ly/EQ8KIrc #IEEE #CircuitsAndSystems #Research #Innovation #IEEEJournals #ElectricalEngineering

17

Learn about "Self-adapting Machine Learning-based Systems via a Probabilistic Model Checking Framework," by Maria Casimiro, Diogo Soares, David Garlan, Luís Rodrigues, Paolo Romano in the 2024 featured articles of ACM TAAS: https://bit.ly/4bgcEEf

2

Word2Vec vs. FastText: Which is better for NLP? Train from scratch or use a pre-trained model? 🤔 I put two NLP giants to the test for semantic analysis in Portuguese!  🚀 The Experiment: 1️⃣ I trained a Word2Vec model from the ground up using the entire Portuguese Wikipedia. 2️⃣ I pitted it against Facebook's massive pre-trained FastText model. 🏆 The Verdict: For general-purpose quality, speed, and robustness, pre-trained FastText is the clear winner! Its biggest advantage? It understands subwords, meaning it can generate meaningful representations even for words it's never seen before. No more "Out-of-Vocabulary" errors! Fun fact: The FastText model was so nuanced it learned that "kitten" (gatinho) is semantically similar not just to "puppy", but also to affectionate terms like "silly one" (bobinho), capturing deep cultural context. A great practical lesson in leveraging state-of-the-art models to accelerate development and achieve superior results. #NLP #DataScience #MachineLearning #Python #Word2Vec #FastText #AI #IBMEC

26

1 Comment

🚀 Borges Full Refactor: Learning AS-to-Organization Mappings with LLMs The Internet isn’t just a graph of Autonomous Systems (ASes)—it’s an ecosystem of organizations constantly reshaping through mergers, acquisitions, and rebrandings. Understanding who runs which networks is critical for management, security, and policy. That’s why we built Borges, a new approach to AS-to-Organization mapping that combines: 📑 PeeringDB metadata (org IDs, notes, aka fields) 🗂️ WHOIS-based AS2Org mappings 🌐 Web signals (redirects, domains, favicons) 🤖 LLM-powered extraction from unstructured text Now, Borges is a full Python package & CLI—complete with tests, CI, docs, and a streamlined pipeline. 🔧 What’s new in the refactor 📦 Modular src/borges/ package with a renewed CLI & new data downloaders 🧪 Full pytest suite + CI across Python 3.9–3.12, style & type checks ⚙️ Stronger WHOIS parsing, consolidation, checkpointing, and logging ⚡ 10× faster batching & rate-limiting in vision-based stages 📚 Expanded docs, hopefully easing reproducibility 💡 Why it matters Borges improves on AS2Org by handling unstructured data, multilingual text, and real-world data—capturing organizational structures at Internet scale. After improving prompt engineering, a full run costs just $3 in OpenAI's API usage. 👉 Try it out and share feedback—especially on edge cases and org-merging heuristics. 🔗 Repo: https://lnkd.in/gG_gRMtH

25

2 Comments

good post, MCP is a good approach and protocol to use different IA systems; i would like to add that we have to mitigate namespace tyosquatting, preference manipulation attack, tool poisoning, rug pulls, coss-server shadowing attack, it must be strengthen code integrity verification. All dependencies and server binaries should be build and verified thought reproducible builds, cryptographic signing and checksum validation to confirm that the deployed code matches audited sources.

1

MLflow Ambassador Joana Ferreira just published an excellent article on building consistent evaluation pipelines for GenAI applications, and it tackles a problem most teams don't realize they have until it's too late. The core issue the article addresses is evaluation drift: when your development testing uses different standards than your production monitoring. You end up with a model that passes all your local tests but behaves differently in production, with no consistent way to compare quality across environments. Joana's solution centers on the "write once, use everywhere" principle: define your evaluation scorers once in MLflow, then apply those exact same scoring functions during local development AND for continuous production monitoring. No more wondering if your prod quality metrics actually match what you tested during development. What makes this walkthrough particularly valuable is the practical implementation. It shows you how to build custom scorers that check both output format and correctness, then demonstrates how the same evaluation logic works seamlessly across your entire pipeline. This kind of systematic approach to evaluation consistency is something more GenAI teams need to adopt. Check out the full article: https://lnkd.in/gC9rSehY #MLOps #GenAI #LLMOps #MLflow #AI

20

1 Comment

https://lnkd.in/eRs2wG6x Solving the kidney exchange problem via graph neural networks with no supervision Pedro Foletto Pimenta Pedro Henrique da Costa Avelar This paper introduces a new learning-based approach for approximately solving the Kidney-Exchange Problem (KEP), an NP-hard problem on graphs. The KEP consists of, given a pool of kidney donors and patients waiting for kidney donations, optimally selecting a set of donations to optimize the quantity and quality of transplants performed while respecting a set of constraints about the arrangement of these donations. The proposed technique consists of two major steps: the first is a Graph Neural Network (GNN) trained without supervision; the second is a deterministic non-learned search heuristic that uses the output of the GNN to find a valid solution. To allow for comparisons, we also implemented and tested an exact solution method using integer programming, two greedy search heuristics without the machine learning module, and the GNN alone without a heuristic. We analyze and compare the methods and conclude that the learning-based two-stage approach is the best solution quality, outputting approximate solutions on average 1.1 times more valuable than the ones from the deterministic heuristic alone.

37

1 Comment

Transformer attention layers operate by projecting high-dimensional token representations into low-rank subspaces defined by per-head key/query/value projections. This inherently places a rank constraint on how relationships between items can be represented, and it forces the model to approximate rich interactions via combinations of rank-limited components. Our paper “Orthogonalization of data via Gromov-Wasserstein type feedback for clustering and visualization” provides a theoretical and algorithmic framework for dealing with full and low-rank structure in relational data representations by adapting transition probabilities and iteratively refining cluster orthogonality via Gromov-Wasserstein-inspired feedback. It shows how a low-rank Markov transition representation can be refined to better reflect true structure and interpretability, and how such refinements converge to meaningful solutions with spectral gap enhancement. This connects directly to issues in multi-head attention: -Both involve representing pairwise relationships (attention scores vs. transition affinities) in a full-rank or approximately low-rank form. -Low-rank factorization can produce spurious grouping or similarity structure that doesn’t reflect the true geometry, unless guided by careful feedback or regularization. -The GW-feedback orthogonalization that we exemplify with shows some ideas how to improve, stabilize, and interpret the geometry of such low-rank relational structures, offering insight into how optimization can go astray in ill-conditioned, low-rank parameter spaces, but also how the implication of the approximation can be assessed. Give it a read: Low-rank constraints are not just computational conveniences — they have real geometric and optimization implications, and without appropriate mechanisms (feedback, orthogonalization, spectral refinement), learned low-rank representations (like attention maps or transition matrices) can be misleading or unstable. I’d like to hear your thoughts on this. Is it needed? Should we care? What’s a good future architecture? What should I read? Merry Christmas!

10

New preprint from PhD student Chenglei Hu in which we apply approximate Bayesian inference methods and extreme value theory, combined with ML feature selection methods, to forecast occurrence and severity of wildfires across Portugal https://lnkd.in/eiTghyx4

12

If you want to build a multimodal RAG application, check the 𝐁𝐮𝐢𝐥𝐝𝐢𝐧𝐠 𝐌𝐮𝐥𝐭𝐢𝐦𝐨𝐝𝐚𝐥 𝐑𝐀𝐆 𝐀𝐩𝐩𝐥𝐢𝐜𝐚𝐭𝐢𝐨𝐧 series of articles I have recently published on To Data & Beyond: 1️⃣ Introduction to Multimodal RAG Applications https://lnkd.in/ddkMHA84 2️⃣ Multimodal Embeddings https://lnkd.in/dYa5PC-x 3️⃣ Multimodal RAG Application Architecture https://lnkd.in/dHxjqUk3 4️⃣ Processing Videos for Multimodal RAG https://lnkd.in/dFfGj-FN 5️⃣ Multimodal Retrieval from Vector Stores https://lnkd.in/daw9d4qP 6️⃣ Large Vision Language Models (LVLMs) https://lnkd.in/dvEVwrc4 7️⃣ Multimodal RAG with Multimodal LangChain https://lnkd.in/djYPWcmY 8️⃣ Putting it All Together! Building a Multimodal RAG Application https://lnkd.in/dGhCdJGV --------------------------------------------- ➡ Do not miss the 50% discount on the To Data & Beyond yearly subscription: https://lnkd.in/dZ2Sjkhk ➡ All My 7 Books, One Button Away, With 40% Off: https://lnkd.in/dqzXcDyJ

35

🔬 I just evaluated 11 retrieval strategies and the results shocked me. Using the Ragas framework, I explored how different retrieval methods perform in real-world RAG systems. Here’s what I found: 1️⃣ Old tech can still win. BM25 from the 1970s! was the fastest retriever (3.08s/query). Sometimes simplicity wins over sophistication, especially when speed matters more than perfect accuracy. 2️⃣ Recall > Precision. The top retrievers hit 100% recall. Missing info kills answers faster than noisy context. Your LLM can filter noise, but it can't invent missing facts. 3️⃣ Measure trade-offs, not just scores. Contextual Compression hit 100% precision/recall but cost $0.056/query vs BM25’s $0.054. At 1M queries, that is $2,300 difference. Production decisions require cost + latency + accurate data. 💭 Open Research Question: Are our synthetic evaluation datasets too friendly to the LLMs that generated them? What is the optimal synthetic-to-real ratio for reliable RAG benchmarks? 📚 Notebook: https://lnkd.in/g-uhq8Rd 🎥 Loom Video: https://lnkd.in/gmBVqwXC What's your experience with RAG evaluation? Have you found synthetic datasets reliable, or do you always validate with real user queries? Huge thanks to AI Makerspace, 👨‍🏫🤖 "Dr. Greg" Loughnane, Chris "The Wiz 🪄" Alexiuk, Laura Funderburk, Ovo Okpubuluku, Tyler Laughlin for inspiring this kind of thinking! #RAG #LangChain #Ragas #LLM #AIMakerspace #AIResearch #AIEngineering

21

4 Comments

This weekend we merged two PyTensor Example notebooks by Ricardo Vieira and Jesse Grabowski - Intro to PyTensor: https://lnkd.in/dkEk7Frr - Normalizing Flows in PyTensor: https://lnkd.in/dzvPEE6k What is PyTensor? A library to define, manipulate, and compile computational graphs. This library is the backbone of PyMC but it has much more potential on its own beyond Bayesian modeling. Check it out!

66

1 Comment

🔥 FineTranslations just dropped: 1 Trillion tokens of parallel text across 500+ languages! Created by Guilherme Penedo and team at Hugging Face, this dataset tackles a major gap: most models struggle with English→X translation, especially for low-resource languages. 💡 The insight: Instead of translating English content into other languages, they started with authentic non-English web pages from FineWeb2 and translated TO English. This creates the perfect parallel data for training better English→X models. 🎯 Key features: - 500+ languages with document-level translations - Smart chunking preserves context across sentences - Fully open & reproducible (ODC-By license) Use it for translation model fine-tuning, multilingual LLMs, or even English pre-training.

53

4 Comments

EASE 2025 Paper spotlight 🔦 (Full research paper) "Assessing the Bug-Proneness of Refactored Code: A Longitudinal Multi-Project Study" By Isabella Ferreira, Lawrence Arkoh, Anderson Uchôa, PhD, Ana Carla Bibiano, Alessandro Garcia & Wesley K. G. Assunção Refactoring: Fix it or break it? Developers refactor code to make it cleaner, easier to maintain, and — in theory — less buggy. But what if some refactorings are actually introducing bugs? This #EASE2025 paper dives deep into 12 popular open-source projects, analyzing over 27,000 refactorings and nearly 55,000 bugs to answer a tough question: 👉 Which types of refactoring lead to more bugs — and how soon do they strike? Key takeaways: a) Single refactorings (like renaming or extracting one method) are over 3x more likely to introduce bugs than composite ones (multiple coordinated changes). b) On average, it takes just 7 changes after a single refactor for a bug to appear. With composite refactors, it takes closer to 29. c) Even well-meaning “pure” refactorings (root-canal style) — meant just to clean up code — can still introduce bugs, especially when done alone. d) Extract Method and Inline Method were the riskiest refactoring types — often leading to bugs despite their simplicity. e)The presence of multiple code smells before refactoring significantly increases the risk of bugs later. Why this matters: a) Teams relying on frequent micro-refactors should favor composite changes and watch for silent regressions. b) Tools could be built to predict bug-prone refactors based on historical patterns. c) Code reviews and CI might need extra scrutiny on certain refactoring types (yes, even the “simple” ones!). For full access to the paper:  https://lnkd.in/dEVVP9b4 🔗 Discover more gems like this at #EASE2025 → https://lnkd.in/duhcSfUp running from June 17-20 2025 Have you ever introduced a bug while “just” cleaning up code? 👇 Drop your refactoring stories in the comments — lessons welcome. #EASE2025 #Refactoring #SoftwareMaintenance #BugProneCode #CodeSmells #DeveloperExperience #EmpiricalSE #SoftwareEngineering #TechDebt #DevTools #SoftwareQuality #SoftwareReliability #QualityAssurance (QA) #CodeQuality #SoftwareEvolution #CodeMetrics #SoftwareAnalytics #LongitudinalStudy #EmpiricalResearch (reinforces #EmpiricalSE) #AcademicResearch #CleanCode #ContinuousImprovement #SoftwareArchitecture #OpenSourceSoftware (if not already included)

12