In a fascinating New York Times Magazine article, Cliff Kuang explores why we as humans can’t trust artificial intelligence (AI) that cannot explain itself. It’s a growing problem, particularly as the European Union’s General Data Protection Regulation (GDPR) is set to go into effect in May. Now a TM Forum Catalyst proof-of-concept project is addressing this issue of AI trust in telecom processes.
The problem is that AI systems work by using probability and correlation, and can arrive at a decision or take an action without humans understanding exactly what’s happened ‘inside the black box’ to achieve the outcome. This lack of transparency is behind a push toward explainable AI (XAI).
“[The goal of XAI] is to make machines able to account for the things they learn, in ways that we can understand,” Kuang writes.
He goes on to explain why this isn’t just a theoretical concern. Two provisions of the GDPR, Articles 21 and 22, address machine learning. Article 21 lets people opt out of personalized ads, while Article 22 gives EU citizens the right to explanation, meaning they can contest decisions made by algorithms and ask for human intervention. Basically, the GDPR rules demand that companies be able to explain to their customers how machine decisions are made.
Making smart BPM smarter
A continuing TM Forum Catalyst project called Artificial intelligence makes smart BPM smarter is looking at how to incorporate AI-based decision modeling and XAI into telecom business processes such as provisioning, fault management, assurance and customer management. The team is using standard decision modeling and notation (DMN) to create a layer of XAI, or what amounts to an AI support system.
The first phases of the Catalyst explained the concept of ‘smart’ business process management (this includes automated business process discovery, validation, mapping and guidance), and developed a prototype communications service providers (CSPs) could use to automatically discover, validate and map business processes to the TM Forum Business Process Framework.
Then at TM Forum Live! last year, the team added AI and software-defined networking (SDN) to show how to discover and repair faults, and train the systems involved to ‘learn’ from experience so that they can automatically take the right action the next time a similar fault occurs.
Developing an AI support system
AT TM Forum Live! Asia in December, the team which included KDDI Research, Orange and Sri Lanka Telecom as champions, and Hewlett Packard Enterprise (HPE) and Trisotech as participants, added the DMN standard as a “wrapper around AI” to make it easy for staff in a telco business office to make and leverage decisions instead of relying on IT staff to do coding, says Ron Ambuter, Senior VP, Trisotech, and leader of the project.
“The idea was to take work that was done in AI and expose it in a format they could understand,” he explains.
In Singapore, the team demonstrated a 5G fault management use case, showing how a telco operations team could train management systems through use of an AI support system. HPE and KDDI have been working together to build the AI system, while Trisotech provides the modeling and coverts the output to executable APIs, which can be contributed back to the Forum. The champions supply the use cases.
Here’s how it works
- The AI support system collects data from other support systems, such as alarm and network management systems, indicating there is a fault.
- The operator checks this information and manually starts the healing process.
- Then the user initiates a process to train the AI system so that it can heal the fault automatically if it happens again.
This process includes looking at whether the repair was successful, the role of the user and their level of skill (whether they’re an expert, for example). Based on this data, an overall score is determined and a recommendation made about whether to include it as training data for the AI system. This way, a CSP can demonstrate the steps taken to automatically resolve a performance issue in a way that a human being can understand.
“Today there is no way to determine that decisions made or processes chosen were the correct ones,” Ambuter says. “Regulatory agencies and courts will ask: ‘How did you come to that decision?’ Saying that an app made the decision for us won’t be an effective answer. We need to understand in human terms what the AI is doing and why. We have a first step here.”
In the next iteration of the Catalyst to be demonstrated at Digital Transformation World in Nice in May, the team hopes to use the same approach to address other telco business processes, such as customer experience management, supply chain management, service fulfillment and service assurance.
“Our goal is to issue a challenge to attendees,” Ambuter says. “We’re willing to help service providers make use of the Catalyst and implement a proof of concept in their company to demonstrate that this isn’t just wonderful abstract thinking.”
He adds that Trisotech and HPE would like to conduct three-month proofs of concept for interested CSPs, during which time they will show operations teams how to use the AI support system to address their specific use cases.
The Catalyst team is looking for additional participants for this phase of the project. In particular they would like to find a company willing to provide a business process management engine or suite. If you’re interested in participating, please contact Tania Fernades, via [email protected].