Learn&Drop: Accelerate CNN Training by Dropping Layers
Learn&Drop speeds up CNN training by selectively dropping layers, cutting computation without losing accuracy. Ideal for VGG, ResNet on MNIST, CIFAR-10.
Enhancing Generative Retrieval: Testing Look-Ahead Prior Robustness
Explore stress-testing and reproduction of the Look-Ahead Prior in generative retrieval, focusing on decoding challenges and cross-lingual robustness.
Parametric Memory Head Boosts Continual Generative Retrieval
Discover how a parametric memory head with PAMT improves continual generative retrieval by balancing stability and plasticity in dynamic document collectio...
Resolving Client Disagreements in Federated Learning Models
Explore a taxonomy and strategy to manage client-level disagreements in federated learning, ensuring fair, efficient model training across diverse clients.
Sinkhorn with Memory for Nonlinear Schrödinger Bridge Control
Discover the novel Sinkhorn recursion with memory to solve nonlinear Schrödinger bridge problems with input and noise channel mismatch.