I had the opportunity to deliver a talk at the iVT Autonomous Industrial Vehicle Technology Conference in Cologne, Germany where I learned about whole new categories of autonomous vehicles for industrial applications.
When most of us hear the term “autonomous vehicle” we probably think of next generation passenger vehicles such as Waymo or Tesla’s electric truck. It turns out that there are whole other classes of autonomous vehicles used for everything from mining to agriculture to driverless forklifts. Autonomous vehicles can help improve safety, boost productivity and efficiency, and reduce cost. In case you think this is the stuff of science fiction, consider that Rio Tinto, a major global mining conglomerate began trials of autonomous industrial vehicles in 2008 and now move roughly one-quarter of all ore and waste material in modern mining operations with driverless trucks.
Teaching computers to drive
Human drivers make decisions such as deciding whether to brake or steer or how to manipulate a robotic implement with relative ease. In an autonomous vehicle, these decisions are based on the level of confidence of a variety of algorithms continually performing tasks such as image recognition, classification, and prediction of where objects will be in the future. Predictive algorithms onboard vehicles are constantly re-evaluating their surroundings, fusing data from on-vehicle sensors, cameras, LIDAR and data other sources.
Even more impressive, the decisions that autonomous vehicles need to make instantly are often the result of multiple predictive models running simultaneously. Ensemble models (obtained by synthesizing the output of multiple analytic models) often deliver better predictive quality than a single model.
By most objective metrics, the control systems that operate these vehicles are already better at operating these vehicles than humans. Walking around the show floor, most conference attendees seemed to be interested in the electronics and control systems that guide these autonomous vehicles. As someone with a background in HPC, I found myself most interested in how we build the AI systems and complex models that enable these vehicles to make decisions like human operators. This is where machine learning and deep learning comes in.
Machine Learning relies on High-Performance Computing
For problems such as computer vision, and recognizing objects in a stream of video data, deep, multi-layer neural networks modeled after the operation of the human brain provide the best predictive accuracy. The challenge with deep neural networks, however, is that they are enormously complex and computationally expensive to train. While the algorithms onboard a vehicle that perform inference (prediction) may execute in a few tens of milliseconds training a single deep neural network model can take days or even weeks requiring large parallel GPU-enabled clusters and massive amounts of training data. Not only does model training involve learning an optimal set of weights for the millions of neurons that comprise a multi-layer neural network, but a separate process called meta-learning involves finding an optimal model architecture, including the number and types of layers, and parameters for the model itself.
HPC has always been an important technology for vehicle manufacturers and computer-aided engineering a major application for workload management software such as Univa Grid Engine. Computer simulation is used for everything from structural analysis to computational fluid dynamics to crash simulation.
Machine Learning is on the rise
At Univa we’ve seen a significant shift in the workloads that industrial manufacturers are running on Grid Engine clusters. Customers are increasingly running machine learning workflows and big data environments alongside existing engineering applications. In fact, in our annual survey of more than 340 technology and IT professionals, 96% of respondents stated that ML projects would grow over the next two years and 69% of companies have three or more teams requesting ML projects.
Machine learning pipelines involve ingesting large amounts of data from various sources including telemetry streaming from autonomous vehicles. Data needs to be cleansed and transformed to create data sets suitable for model training. These data sets are then used with machine learning frameworks such as Google’s TensorFlow and other frameworks to create, train and validate predictive models.
In some cases, predictive algorithms are forward deployed in silicon or in software on the autonomous vehicle, and in other cases, predictive models are deployed as inference services on the cluster where they need to be managed and prioritized along with other engineering and machine learning workloads. Also, model training never stops. Engineers need to continually tune and validate models based on new datasets as vehicles find themselves operating in different conditions or used for different applications.
From an HPC perspective, machine learning workloads exhibit many of the same challenges as other applications. The workload manager needs to manage multiple workload requests with complex resource requirements from multiple users and schedule them optimally to maximize productivity and resource utilization while ensuring deadlines are met.
Grid Engine – A better platform for Machine Learning
Machine learning and deep learning workloads pose some unique challenges. At Univa, we’ve been busy delivering innovations in Grid Engine aimed at making it the industry’s best workload manager for distributed machine learning applications. Some of these innovations are listed below:
Based on the strength of its GPU management and machine-learning friendly features, Univa Grid Engine has become a preferred platform for some of the world’s leading AI supercomputers. Univa Grid Engine today powers Japan’s RAIDEN (Riken AIP Deep learning ENvironment) supercomputer and the AI Bridging Cloud Infrastructure, #5 on June 2018 ranking of the world’s top 500 supercomputers.
Combining HPC and AI for next-generation autonomous vehicles
The future of autonomous industrial vehicles is bright. As usage increases, and we find more industrial applications for autonomous vehicles, we will need ever more advanced control systems powered by increasingly sophisticated AI and Deep Learning algorithms. Are you running machine learning or deep learning workloads on your Univa Grid Engine cluster? If so contact us. We’d love to learn about your experiences!