Back

Speaker "Hagay Lupesko" Details Back

 

Topic

Title: Deep learning in production: going beyond Python
Title: Deep learning acceleration with Amazon Elastic Inference

Abstract

Title: Deep learning in production: going beyond Python
Abstract:
Deploying deep learning models in production requires various considerations, including: security, performance, scalability, and integration into existing production systems. While Python is the most popular programming language to build and train deep learning models, often it will not be the best choice for addressing the considerations mentioned above and using models in production. In this talk, we will go beyond Python, explore language APIs for inference across popular languages such as Java and C++, and discuss the benefits, as well as the technical risks, involved in using these APIs for deep learning inference in production.


Title: Deep learning acceleration with Amazon Elastic Inference
Abstract:
Deep neural network inference workloads typically require GPU acceleration to achieve low latency and high throughput. While modern GPU compute is  relatively expensive, most workloads do not utilize the entire compute capacity of the GPU. In this talk, we will dive into the internals of deep learning inference on GPU, and how you can use Amazon Elastic Inference to attach the GPU compute capacity needed by your model, and achieve up to 75% reduction in inference cost. By the end of this talk, you will learn how you can get started using Amazon Elastic Inference for your deep learning inference workloads.



 

Profile

Hagay Lupesko is part of the deep learning leadership team at Amazon Web Services, and currently works to democratize Artificial Intelligence and Deep Learning through cloud services and open source projects such as MXNet and ONNX. He has been busy building software for the past 15 years, and still enjoys every bit of it (literally)! He engineered and shipped products across various domains: from 3D cardiac imaging with real time in-vessel tracking, through semi-conductors fab systems that measures structures the size of molecules, and up to web-scale systems with global distribution.