Even as machine learning (ML) algorithms become more sophisticated and powerful, the way ML teams build ML systems hasn’t changed much. In this article, we’ll explain the need for composable machine learning systems. First, take a look at the old, inefficient way. Once the team figures out the task (e.g. classify data, detect regions of interest, forecast the future, find anomalies, etc.) and the type of data involved (is it tabular, time-series, graph, image, natural text, or other?), the usual next step is to design and program a mathematical model and computational algorithm from scratch, or search for and download code for a matching pre-built model/algorithm. The team prepares the data pipeline, trains or fine-tune the model using the algorithm, validates, iterates, deploys, serves the ultimate solution, and upgrades to a better model when that becomes available.
But what happens when the task is highly complex, such as controlling an industrial manufacturing process (such as cement, steel production or oil & gas extraction) in real-time to improve its yield, power and raw material consumption? Or writing an x-ray medical report on behalf of a doctor? Or a cognitive chatbot that answers questions on behalf of customer support? Or when the task needs two or three data sources, such as radiological images plus doctors’ notes? Chances are there won’t be a ready-to-download model that nicely fits these requirements, so the team needs to build their own unique and substantially different model. Realistically, the great majority of ML teams find themselves limited to using downloadable models, and thus unable to innovate. Some teams might have the skills and research training to go farther and produce a hand-crafted, one-off solution—only to find that it cannot be scaled into production.
To break this status quo, old systems need to turn into composable machine learning, so that ML teams—not necessarily made up of advanced experts—can build applications for a richer spread of AI tasks, and take them into scalable production. Composable ML is not the same as ML research, which can be thought of as creating new model and algorithm building blocks. Instead, composable ML makes it simpler to design ML models—like electricians building or installing circuits at many different houses—by allowing non-researchers to put existing building blocks together, or to recombine them to solve new tasks.
Texar is one such composable machine learning development tool developed at Petuum, that shortens the ML development cycle by allowing ML teams to assemble complex ML systems in a symbolic manner. Our initial release of Texar is focused on Natural Language Processing, and we plan to expand it to other AI fields such as Computer Vision and Time Series modeling. Texar decomposes and represents complex ML systems via their architectural elements—such as mathematical models, loss functions, constraints, learning and inference algorithms, or more complex compound ML modules such as encoders and decoders, and more.
At Petuum, we use Texar to create sophisticated and original ML systems for medical report writing from chest x-ray images, and multi-lingual (English, Chinese and Japanese) cognitive chatbots for retail in-store assistance and call center support, as well as reproduce and extend recent models from the research community such as BERT. We provide the Texar-created ML compositions with scalable ML infrastructure support that handles raw data ingestion and pre-processing, distributed model parameter training, elastic resource management, model versioning, training, and inference serving, as well as containerized program management. We believe that all these engineering elements beyond just ML models and algorithms are needed to create non-trivial ML systems that generate enormous customer value, yet do not burden customers with complex data and ML maintenance requirements that are un-scalable. Petuum offers Texar as open source under a friendly license, and we hope that it can benefit other ML teams searching for a sustainable way to produce the next generation of AI applications.
Eric P. Xing
Eric P. Xing is a Professor of Computer Science at Carnegie Mellon University, and the Founder, CEO, and Chief Scientist of Petuum Inc., a 2018 World Economic Forum Technology Pioneer company that builds standardized artificial intelligence development platform and operating system for broad and general industrial AI applications. He completed his undergraduate study at Tsinghua University, and holds a PhD in Molecular Biology and Biochemistry from the State University of New Jersey, and a PhD in Computer Science from the University of California, Berkeley. His main research interests are the development of machine learning and statistical methodology, and large-scale computational system and architectures, for solving problems involving automated learning, reasoning, and decision-making in high-dimensional, multimodal, and dynamic possible worlds in artificial, biological, and social systems. Prof. Xing currently serves or has served the following roles: associate editor of the Journal of the American Statistical Association (JASA), Annals of Applied Statistics (AOAS), IEEE Journal of Pattern Analysis and Machine Intelligence (PAMI) and the PLoS Journal of Computational Biology; action editor of the Machine Learning Journal (MLJ) and Journal of Machine Learning Research (JMLR); member of the United States Department of Defense Advanced Research Projects Agency (DARPA) Information Science and Technology (ISAT) advisory group. He is a recipient of the Carnegie Science Award, National Science Foundation (NSF) Career Award, the Alfred P. Sloan Research Fellowship in Computer Science, the United States Air Force Office of Scientific Research Young Investigator Award, the IBM Open Collaborative Research Faculty Award, as well as several best paper awards. Prof Xing is a board member of the International Machine Learning Society; he has served as the Program Chair (2014) and General Chair (2019) of the International Conference of Machine Learning (ICML); he is also the Associate Department Head of the Machine Learning Department, founding director of the Center for Machine Learning and Health at Carnegie Mellon University; and he is a Fellow of the Association of Advancement of Artificial Intelligence (AAAI), and an IEEE Fellow.