This is the first in a series of interviews with the people who power Motion Metrics. We’re a diverse and tight-knit team of innovators – and although we love downing waffles and talking trash at the foosball table, our real passion lies in driving industry change through leading-edge technologies. Today, we sat down with Mahdi Ramezani, our Director of Core Research and Development, to talk about the ways that artificial intelligence is transforming our company.
What brought you to Motion Metrics?
I became involved with Motion Metrics towards the end of my Ph.D., when my supervisor introduced me to the company’s CEO and Founder, Shahram Tafazoli. Shahram wanted to build a deep learning team within Motion Metrics and incorporate artificial intelligence into the company’s product portfolio, and I had spent the last few years developing deep learning techniques for medical image analysis. It was a great fit. I began as a consultant, and then joined Motion Metrics full time.
Why was Motion Metrics such a great candidate for deep learning?
When I joined Motion Metrics, the company had been in business for more than 10 years and had captured vast datasets, including hours of video footage from mine sites around the world. As a researcher in the field of machine learning, this was a dream – we could take advantage of all that data!
What were the main barriers to implementing deep learning at Motion Metrics?
There were three main challenges. Firstly, it was a difficult buy-in process. We install solutions on many different types of equipment in some of the harshest and most diverse environments around the world. Before using a deep learning approach, Motion Metrics had used hand-coded heuristic image processing algorithms to account for each variation in these operating environments. Internally, there was some doubt that a deep learning model could be sufficiently generalizable to be deployed at our existing mine sites – let alone those we had yet to encounter.
Secondly, our data – which, in my opinion, is our most important company asset – was not organized. We first had to organize all our data based on the mine, equipment type, environmental conditions, application, etc., and then manually label it, which required a big effort.
Computing resources and hardware were also a challenge since we needed to process video frames in real time on existing Motion Metrics embedded devices, which are designed to withstand the harsh mining environment. To make accurate predictions, deep neural networks need a lot of computing power. Most artificial intelligence companies do almost everything on the cloud to take advantage of ‘unlimited’ computing power. Due to limited internet connectivity on mining sites and our strict timing requirements for mission critical applications, we could not rely on cloud computing. Instead, we had to optimize state-of-the-art technologies to overcome this challenge.
How has deep learning benefited Motion Metrics?
Deep learning has helped us improve the accuracy of our solutions, which ultimately makes our customers happier. Take ShovelMetrics™, for example. In the past four or five years, we haven’t introduced any new features – instead, we’ve put all our focus into making our products the best that they can be. And now we’re seeing the pay-off reflected in customer feedback and repeat orders.
What is the current state of the industry? And what does the future hold for artificial intelligence?
The field is changing so rapidly that it’s hard to keep up. Right now, systems using artificial intelligence match or surpass human-level performance in many domains, but those successes haven’t translated into productivity gains.
One possible reason for these disappointing growth rates is that artificial intelligence has been a little overhyped. This often happens when a new technology becomes a public matter, and everyone talks about it. The wild predictions have now subsided, but that doesn’t mean that the technology has stopped progressing – just that it’s being developed and deployed at a more realistic pace. Another possible cause could be the implementation lag in using artificial intelligence to its full potential.
Most developed algorithms today can be run by anyone. Even a high school student can use some of the open-source developed libraries to run a deep neural network for a task like object detection. But there’s a difference between running a program and knowing how it works, how it can be improved, and whether the approach fits the problem. To drive business value, companies need subject matter experts with a strong theoretical background and experience in software development.
In what direction will artificial intelligence take Motion Metrics?
There is a lot of room to expand our use of artificial intelligence at Motion Metrics. In the future, we can automate many mining workflows to increase efficiency.
Take our missing tooth detection solutions, for example. Currently, we notify the operator that a tooth is missing using the latest in artificial intelligence, then the operator manually notifies dispatch, and then finally dispatch brings a replacement tooth. Right now, the solution provides information about the critical event but doesn’t use that data to take action. Future control systems could stop the shovel from digging as soon as a missing tooth is detected while simultaneously arranging for a tooth replacement, rerouting the truck carrying the missing tooth, and optimizing truck dispatch to account for the downtime.
Mahdi Ramezani has been with Motion Metrics since 2014 and is the Director of Core Research and Development. He holds a Ph.D. in Electrical and Computer Engineering from the University of British Columbia in Canada, as well as an M.Sc. and B.Sc. in Electrical Engineering from the Sharif University of Technology in Iran.