After clicking “Watch Now” you will be prompted to login or join.
Neuroevolution-Based Automated Model Building: How to Create Better Models
Keith Moore, SparkCognition
GTC 2020
Common neural network architectures may work well for known, established data problems; however, they can fall short when modern machine learning applications demand more performance and higher levels of sophistication. In this presentation, Keith Moore, director of product management at SparkCognition, covers how neural architecture search works, some of the challenges faced in the space, and why an evolutionary approach is capable of discovering sophisticated and elegant designs that fit your data. He'll discuss how multiple models can be trained in parallel using MPS technology, how batch size loading can change performance curves, and how the use of PyTorch and automatic mixed precision on a Tesla V100 GPU can decrease training times by over 5x. In conclusion, Keith will take you through the journey their team faced on building out better models, provide some real-world examples and implementations, and also share some of the problems that are yet to be solved.