My Notes From Spark+AI Summit 2020 (Application-Agnostic Talks)

Eugene Yan · 6 years ago

The Spark+AI Summit 2020 conference in June featured technical talks on scaling deep learning models through efficiency techniques like pruning, quantization, and distillation, as well as practical optimization methods for distributed training and data processing. Key concrete examples included MobileNet V1 reducing parameters from 24 million to 4 million with 78.8% to 70.9% accuracy tradeoff, and pruning achieving 50% weight reduction with minimal accuracy loss on ImageNet. Organizations can now deploy smaller, more efficient models in production using frameworks like PyTorch with integrated tools for serving, distributed training, and hardware acceleration.