Lowering the Barriers to Telco AI With Acumos
James Crawshaw, Senior Analyst Service Provider IT and Automation, Heavy Reading
The Linux Foundation's Acumos AI Day at the Orange Developer Center in Paris on July 13 showed how artificial intelligence capabilities are already delivering benefits to network operators, albeit in a rudimentary way.
The event was hosted by Orange (NYSE: FTE) executives who are responsible for coordinating the various open source activities in which the French incumbent is involved, including ONAP, OpenStack, OPNFV, OpenDaylight (ODL) and, more recently, the Acumos AI Project. There was also representation from Deutsche Telekom AG (NYSE: DT), which already has many AI projects ongoing across the business and sees Acumos as a way of bringing these closer together with a common framework.
Several vendors were present, some of which were keen to pursue a Red Hat-like business model with Acumos, some of which were looking to get their own IP aligned with the project, and some of which were just trying to figure out what it all means.
So, what is Acumos? If you really want to know, I highly recommend the white paper written by AT&T's Reuben Klein. Acumos is essentially an AT&T Inc. (NYSE: T) lab project that has been spun out to the Linux Foundation in order to apply the concept of crowdsourcing to software development. The beta version of Acumos has code from AT&T and Tech Mahindra and is available to download now. The project is hosted by the Linux Foundation's Deep Learning group, the founding members of which are (in alphabetical order): Amdocs, AT&T, BYond, Baidu, Huawei, Intel, Nokia, Orange, Red Hat, Tech Mahindra, Tencent, Univa and ZTE. (See Linux Foundation Injects Open Source Into AI and LF Networking Gets Ambitious About Open Harmonization.)
In the white paper, Klein explains how the development and deployment of AI applications is currently highly time-consuming and requires expensive, specialist talent. Acumos will provide a common framework that reduces the need for machine learning (ML) "rocket scientists" and accelerates development, thereby lowering the barriers to AI for telcos and companies in other industries too.
With Acumos, general data scientists (who are not also ML specialists) will be able to create AI models built from tools such as Google's TensorFlow, Scikit-learn, R-Cloud or H2O (or directly programmed in Java, Python or R). Acumos will then package the model into a containerized microservice so that it can be easily integrated into applications by ordinary software developers without the need for a masters degree in data science and a specialism in AI development tools.
What makes Acumos even more interesting is that it borrows from the app store concept by providing a marketplace where academics and commercial data scientists can try to sell specialized machine learning models, customized to address a specific industrial use case. Telcos (and other companies) can then experiment with those crowdsourced models on their own datasets and, if they prove useful, pay to use them in production.
Acumos has taken the shrewd decision to leave data storage and processing outside its remit so that it is not exposed to any data privacy risk issues. Telcos will need to use their existing big data platforms (Kafka, Hadoop, Spark, map reduce, NoSQL, etc.) and feed their chosen models with this data within their own environment.
So, is this just a build-it-and-they-might-come type of science project? Well, based on the discussion at the meeting, I think that AI applications are still quite nascent, even among the early-adopter telcos that were present. However, AI is already demonstrating benefits in areas such as alarm filtering. The specific areas of investigation discussed in the white paper include:
- Networking -- responding to traffic anomalies in the real time (e.g. spinning up servers to meet demand spikes)
- Infrastructure -- more accurate capacity planning
- Security -- learning new attack signatures on the fly
- Customer care -- (more) intelligent chatbots
The first release of Acumos (called Athena) is due in November, by which time hopefully there will be some more apps in the marketplace. Currently there are just four from AT&T (image classification, face detection, face privacy filtering and image mood classification) and two from Tech Mahindra (cross selling and customer segmentation). The current paucity of apps/models is partly because the monetization details (licensing) haven't yet been implemented and because there is not yet an automated mechanism for scanning models for malware.
Even if these issues are not resolved by November, I would hope that AT&T could gift a few more models to help kickstart the project. Acumos today is a small acorn, but given time and a bit more operator and vendor love, it could grow to become a key enabler of AI-driven automaton in the telecom sector and beyond.
James Crawshaw, Senior Analyst, Heavy Reading