[105th TrustML Young Scientist Seminar] Talk by Prof. Kfir Levy (Technion) "Stable Optimization for Robust Federated Learning: A Double Momentum Approach"

The 105th installment of the TrustML Young Scientist Seminar features Prof. Kfir Levy of the Technion, who will present work on stable optimization for robust federated learning. The talk runs for one hour on the morning of July 30th, 2026, and is delivered online, with an on-site open space reserved for AIP researchers. Levy's talk introduces an optimization method that reaches the convergence rates of SGD while keeping the stability and simplicity of full-batch gradient descent. The method builds a gradient estimate out of two recent momentum mechanisms, which lets it run at a fixed learning rate without the usual learning-rate tuning or a held-out validation set for monitoring generalization error. The second half of the session turns to distributed settings, where the double momentum variant yields tighter theoretical bounds for differentially private and Byzantine federated learning, with personalized federated learning covered if time allows. Levy is an associate professor of electrical and computer engineering at the Technion working on machine learning, AI, and optimization, and holds the Alon, ETH Zurich, and Irwin & Joan Jacobs fellowships.