Accelerating ELMs on the GPU Toward Real-Time Training on Large Scale Data Sets