Journal of Artificial Intelligence Research

Excited to start as a co-editor-in-chief at the Journal of Artificial Intelligence Research (JAIR). Some of my own favorite papers were published at JAIR, and the journal has consistently published high-quality papers over the last decades and continues to do so among all the AI turbulence going on nowadays, so it is a privilege to be able to contribute back! I am sure there will be moments that I might regret having signed up to deal with the tsunami of papers, but I trust that, together with this great team, we'll be able to handle it :) #AI

I'm excited to share that our paper, "Deep Neural Network Compression via Data-Driven Low-Rank Singular Value Decomposition," has been published in the Journal of Artificial Intelligence Research (JAIR). In this work, we introduce a novel data-driven approach to SVD-based neural network compression. Rather than relying solely on network weights, our method leverages representative input data to better preserve the information that matters during compression. This leads to significantly improved performance over existing SVD-based techniques while substantially reducing model size and computational cost. https://lnkd.in/esf3trMk #ArtificialIntelligence #DeepLearning #MachineLearning #NeuralNetworks #ModelCompression #Optimization #SVD #Research #JAIR

🎉 Excited to share that this is our second paper published in the Journal of Artificial Intelligence Research (JAIR)! 📄 Formal Logic Inference Guided Uncertainty Quantification for Personalized Federated Learning https://lnkd.in/grf_Vda7 Federated Learning (FL) has emerged as a powerful paradigm for training machine learning models while preserving data privacy. Yet, real-world deployments—such as smart grids, traffic forecasting, and IoT systems—must contend with highly heterogeneous client data, making personalization both challenging and computationally expensive. In this work, we introduce LogiCP, a novel framework that combines formal logic reasoning with uncertainty quantification to enable scalable and personalized federated learning with theoretical guarantees. Our key idea is to leverage Signal Temporal Logic (STL) to identify meaningful temporal behaviors and group clients based on semantic similarity rather than purely statistical distance. Within each cluster, we apply decentralized Conformal Prediction (CP) to generate distribution-free prediction intervals with mathematical coverage guarantees. LogiCP also supports dynamic client assignment, allowing new clients to join appropriate clusters at runtime without retraining. Across three real-world applications—traffic forecasting, temperature prediction, and electricity demand forecasting—LogiCP consistently outperforms Bayesian, clustering-based, and conformal prediction baselines, achieving up to a 95% improvement in client-level MSE while maintaining strong scalability. More broadly, this work demonstrates how formal methods can play a central role in modern AI—not only for verification, but also for guiding learning, improving personalization, and quantifying uncertainty. We hope LogiCP contributes toward building AI systems that are more accurate, scalable, and trustworthy for distributed, real-world applications. Many thanks to our co-authors (Guocheng He, Ziyan An), collaborators, reviewers, and everyone who contributed to this work. I'm excited to continue exploring the intersection of formal methods, machine learning, and trustworthy AI. #JAIR #FederatedLearning #MachineLearning #TrustworthyAI #FormalMethods #UncertaintyQuantification #ConformalPrediction #DistributedLearning

🎉 Excited to share that our paper, "Towards Verified and Targeted Explanations through Formal Methods," has been published in the Journal of Artificial Intelligence Research (JAIR)! As AI systems become increasingly integrated into safety-critical applications such as autonomous driving and medical diagnosis, explanations need to be more than intuitive—they need to be trustworthy. In this work, we introduce ViTaX (Verified and Targeted Explanations), a formal explainable AI (XAI) framework that bridges the gap between interpretability and formal verification. Rather than simply explaining a model's prediction or verifying robustness to the nearest decision boundary, ViTaX asks a more meaningful question: How resilient is a model against a specific, high-risk alternative? For example, in autonomous driving, mistaking a Stop sign for a 60 kph sign is far more consequential than confusing it with a No Passing sign. ViTaX generates verified targeted semifactual explanations with mathematical guarantees, certifying that even if the most critical features are perturbed within a specified bound, the model's prediction remains robust against a user-specified target class. Across multiple image classification benchmarks (MNIST, GTSRB, EMNIST) and the TaxiNet regression task, ViTaX achieves substantially higher explanation fidelity—improving over existing approaches by more than 30% while producing concise explanations. This work represents a step toward making AI explanations not only interpretable, but also verifiable, targeted, and trustworthy for safety-critical systems. A huge thank you to my amazing co-authors (David Wang, PhD, Taylor T Johnson, Ipek Oguz), collaborators, reviewers, and everyone who provided feedback throughout this journey. 📄 Read the paper here: https://lnkd.in/gChHFamN #FormalExplainableAI #XAI #FormalMethods #TrustworthyAI #JAIR