#009 - AI-based industrial systems

January 14 2021/30 min read

Our guest today is Nasim Farahini, CTO in AI, IoT and Cloud at Qamcom. She has a rare breadth of knowledge on how to bring AI and other software solutions into the real world, which, as we know, is where the value of these technologies actually is created. 

Key takeaways from this episode

What is an example of an end-to-end system?

  • A multi-function sensor system for road crossings with a camera as the main sensor, where the main decision-making is happening on local server. This model then identifies if and what type of object is at the crossing and sends a control signal to the traffic light.
  • Power line inspection using drone technology
  • Detecting and classifying farm animals using drones for daily surveillance

Are your drone models also edge devices where they are not just collecting data but also having edge technology on the drones?

  • This is a really good example of how end-to-end systems are optimized for cost. If you put the entire algorithm on the drone it will be power-hungry and the drone will run out of power very quickly, but if you also capture all data from drone that will also be unnecessary storage costs. So instead Qamcom puts a light-weight model on the drone that identifies what images are relevant images and only communicates those to local service or cloud, meaning it is optimized both for power and for cost.
How does the hype around AI impact your ability to get things done with customers? 

  • There is still a big knowledge gap around what AI actually is and how it can be used. Sometimes the problems that the customers are presenting are better solved with traditional signal processing, which is much less complicated than using AI. It is important for customers to understand that AI today is just one of the tools in the tool box and that the focus should instead be trying to solve the problem as needed.

What are some of the real-life safety concerns that you’ve had to confront in your work?

  • When the decision that is made has physical consequences, safety is a huge part of the system design but the black-box nature of AI can sometimes make things more challenging. 
  • Qamcom has a large team of safety experts at company. When starting to think about system, they involve safety experts right away to help them put safety into the architecture to avoid it becoming an afterthought. 
  • It is important to design systems for safe failure. When using AI, you have to constantly monitor AI model - e.g. through confidence score to identify instances where AI output is not valid. Typically AI models perform well on data that they have been trained on, but when a new context arises, they might not be as good. It is important to identify when those things are happening and implement statistical models for uncertainty estimations.