Thalaiyasingam Ajanthan

4 papers Australian National University

• Mixtures of Subspaces for Bandwidth Efficient Context Parallel Training
• Sharper Convergence Rates for Nonconvex Optimisation via Reduction Mappings
• Subspace Networks: Scaling Decentralized Training with Communication-Efficient Model Parallelism
• Unextractable Protocol Models: Collaborative Training and Inference without Weight Materialization