Towards New Deep Learning Abstractions on Top of Existing Frameworks
Explore the Nervana Graph (ngraph) Python library for implementing neural network programs in this conference talk from Strange Loop. Discover how ngraph converts neural network descriptions into efficient programs for various platforms. Learn about the library's three-layer structure, including the API for creating computational ngraphs, higher-level frontend APIs for TensorFlow and Neon, and a transformer API for compiling and executing graphs on GPUs and CPUs. Gain insights into the library's design principles, focusing on modularity, flexibility, and computational efficiency across different hardware configurations. Presented by Tristan Webb, an Intel Nervana researcher with a background in computational neuroscience and machine learning, this talk offers valuable perspectives on advancing deep learning abstractions and improving model execution efficiency.