Bengaluru-based Aprecomm provides AI-enabled Wi-Fi solutions for enterprises, SMBs, retail and ISPs. The platform’s AI stack transforms Wi-Fi routers from mere connectivity devices to smart devices.
“At Aprecomm, we are fusing data science with domain expertise to build what we call autonomous networks which are agile and responsive through a collaborative approach, by distributing the intelligence between the edge and the cloud, ensuring optimal resource usage and faster responses in real-time,” said Guharajan Sivakumar, CTO and co-founder of Aprecomm.
Sign up for your weekly dose of what's up in emerging technology.
In an exclusive interview with Analytics India Magazine, Guharajan spoke about Aprecomm’s approach to AI governance.
AIM: What’s Aprecomm trying to solve?
Guharajan Sivakumar: CSPs (Communications Service Providers) are putting tremendous efforts to orchestrate the content to end-users seamlessly through their complex network infrastructure consisting of wifi access points, SD-WANs, switches, routers, distribution switches, BGP routers and MPLS networks. All these elements play their part to meet the needs of customers.
Aprecomm solves for:
· How do we make these networks aware of what customers want and what actions can be taken to meet their needs?
· How well the coordination can be.
· What are the points of failure? How do we predict and avert them?
· How do we maintain these networks seamlessly?
· Can the networks be put in a self-drive mode?
· Can networks be self-driven with minimal manual intervention?
· Can the networks be smart enough to identify the failure and provide the details to the IT Admins and MSPs?
· Can the networks troubleshoot?
· Can they become self-aware?
· Can they self organise?
AIM: Tell us about Aprecomm’s AI governance framework.
Guharajan Sivakumar: Aprecomm collects data from both wired and wireless devices (like phones, tablets, smartwatches, TV, and laptop) in the network. It gathers information about the usage patterns, nature of the applications, radio level characteristics of the devices etc. Based on the raw data collected, QoE of the end customers are measured using supervised, unsupervised AI models, and is compared against the target SLAs. AI models take the decision to adjust wireless parameters or bandwidth limits to ensure the SLA is met for all the end-users connected to the networks. QoE measurement acts as a feedback system to ensure whether the decisions taken by our engine are proceeding in a constructive fashion or if there is a need for manual intervention to override these corrections.
Today’s AI systems are still naive. We employ the following governance mechanisms to keep the decisions more accurate:
· Fairness: Aprecomm employs supervised models to automate certain complex network tasks like throughput estimate, scoping and root cause analysis. It is extremely important to ensure AI engines are not trained with only one type of data. For example, if we train the AI models with more samples of Samsung phones, our models would be able to find more issues with Samsung phones. It is possible that they fail to understand the issues generated by, say, Apple phones. The reports might give an impression that most Samsung phones are problematic and can impact customer buying patterns. Aprecomm avoids these issues with the right mix of training data. Also, Aprecomm has a bunch of periodic manual intervention processes where our machine learning engineers would inspect the models to understand their fairness, particularly with unsupervised models which are learning real-world data. Outliers have to be identified and removed so that AI models won’t get affected by these patterns.
· Awareness: AI engines are black boxes. Aprecomm has built test data which are periodically fed to the models to understand how they are making the decisions. As the behaviour of wifi devices is more dependent on the type of hardware chip used and their components, it is difficult to generalise. This is the reason Aprecomm has built a test set leveraging decades of network experience covering behaviours of various chipsets, internal/external antenna patterns, powers, gains etc.
· Transparency: Aprecomm believes evidential intelligence is key to increasing the acceptance of network intelligence in enterprise and service provider markets. Aprecomm’s E2E Insights provide the network experts with very minute details about how our deployed models are working and offer transparency, and accountability.
· Privacy and security: Aprecomm doesn’t collect any data specific to the customers like names, emails, phone numbers, age etc. Though Aprecomm is focused on QoE of the customers, it is primarily interested in the following non-privacy parameters which can affect experience:
- Amount of time customer devices takes to connect to the internet.
- Amount of data they are downloading.
- Type of applications on the devices.
- Are these applications sensitive to jitter and latency?
- Is the network able to meet the demands of the customers?
- Is there any packet loss in the ISP network?
AIM: What explains the growing conversation around AI ethics, responsibility, and fairness?
Guharajan Sivakumar: It could be used in constructive ways or in destructive ways. AI models are more like newborn children. Just like how children observe adults and start learning what they see and perceive, AI models find patterns in the data and start learning them. If the data has a bias, then the AI models start learning the same bias. This can have detrimental effects based on where the systems are deployed. It could be in places like taking hiring decisions for a company or it could be in defence systems or autonomous vehicles.
Take the example of an autonomous vehicle crash. Who must be held accountable? Is it the customer who has bought this vehicle? Or is it the manufacturer who made this vehicle? Or is it the person who wrote this AI software? How do we even isolate the AI model which has made the wrong decision? What if the AI model has taken a decision to brake, but the brake did not get deployed in time? This is why machines must be transparent enough to provide information about the decision taken.
In the case of networks, what if the model starts tagging a specific home as hampering the business development and starts reducing their bandwidths. It could affect some critical communication resulting in a bigger loss to the customers. There should be a guideline around to what extent this intelligence should be used and where the lines should never be crossed.
To guard privacy, regulatory bodies are enforcing standards like GDPR and PDPB. It’s important these regulatory requirements are followed to ensure AI ethics are met at scale. It is also equally important to come up with guidelines in terms of accountability and transparency of the system.
As we are primarily focused on the networking area and the interaction with the AI engine is with the devices and network packet format, providing the evidence of the decisions taken is incredibly important. Aprecomm firmly believes in evidential AI. Every decision and the conditions resulting in that specific decision is always available to the engineers as well as to the customers for analysis.
AIM: How do you mitigate biases in your AI algorithms?
Guharajan Sivakumar: Bias introduced in AI can lead to catastrophic failures like complete outage of the network. As an AI company, it is important to follow an ethical model to collect data. A scalable feedback system that can take passive and active feedback from users should be deployed to ensure a decision is not biased and the learning is continuous on large scale deployments.
While AI-powered automation fuels the end-to-end experience at Aprecomm, we define a standard set of acceptable metrics and a dashboard to present the key metrics and their variations visually to our customers to evaluate models. This visibility ensures if a bias is present, it gets addressed immediately.