“TensorFlow: A System for Large-Scale Machine Learning” was published on arXiv in May 2016 by Martin Abadi, Jeff Dean, and a large team at Google, and presented at the OSDI 2016 systems conference. It is the technical paper behind the framework Google had open-sourced in late 2015.
The paper describes TensorFlow as a system built around dataflow graphs. A computation is expressed as a graph of operations through which tensors (multidimensional arrays) flow, and the system maps the nodes of that graph onto many devices, CPUs, GPUs, and Google’s then-new TPUs, across a cluster. This design let the same program run on a single phone or across hundreds of machines, and it unified training and inference under one representation. The paper emphasizes flexibility for representing a wide range of algorithms and the mechanisms for distributing state and computation, including how parameters are shared and updated across workers.
TensorFlow’s graph-based, multi-device approach made industrial-scale machine learning broadly accessible and helped set the terms of the framework era that followed.
For a business reader, this paper documents the infrastructure that turned deep learning from a lab technique into something deployable across data centers and devices.