Unmanned Systems Technology 013 | AutonomouStuff Lincoln MKZ | AI systems | Unmanned Underwater Vehicles | Cosworth AG2 UAV twin | AceCore Neo | Maintenance | IDEX 2017 Show report

Teaching a neural network requires high-performance processing. Here, multiple GPU cards are combined to provide 44 TFLOPS of processing (Courtesy of Horiba MIRA) 33 Deep learning takes the video feeds from the cameras, but this data has to be combined with data from other sensors such as Lidar and ultrasound, GNSS navigation data and mapping information. For example, if a Lidar sensor detects a bump in the road, a camera can then determine whether it is a speed bump or someone lying in the road. Alongside this, the vehicle needs to understand the world around it, how other objects such as cars and pedestrians (known in this context as agents) are moving and, vitally, what they are likely to do. This agent-based analysis is an essential part of the structure of an AI system for autonomous vehicles. There is also a fundamental difference in the technology required for different types of applications. For example, an autopilot application on a busy road where the driver is behind the wheel and ready to take over if necessary has different AI requirements from an autonomous vehicle that may hand over to a remote operator, and different requirements again from a fully autonomous design. The performance of roadside sign detection systems is geared to a certain level of performance, of around 75%. That isn’t sufficient for an unsupervised system though, which needs to reach 99.999% accuracy. That is not possible with traditional image recognition algorithms, and this type of requirement is pushing the adoption of systems based on neural networks. Elements of AI Deep learning for image recognition and classification has been the main focus of AI research over the past few years, and there are several types of neural network that can be used here (see sidebar, page 36), depending on the type of images being provided. The neural network itself is a series of computation nodes in several layers. Each node in a layer is connected to each node in the layer below, and the strength of the connections between them is determined by a weighting. A positive weighting reinforces a result, a negative one reduces it. Hundreds of thousands of images, all tagged with information about their content, are fed into a neural network and the output information compared to the original image. In a backpropagation network, this process is repeated, adjusting the weights until the input and output agree. The richer the data, the more accurate the overall process, but marking the images and training the neural net can be an expensive process and take weeks or months, even with a high-performance server. Many of the deep learning algorithms being used for image recognition in autonomous vehicles are using these backpropagation networks, although there is intensive research into progressive networks and unsupervised networks that do not require backpropagation and so can be more efficient. While these are being used in trading and data centre AI systems, the challenge is to improve the accuracy of these newer networks for image recognition in autonomous vehicles. The training can be performed in the cloud or on dedicated servers in the lab. One training computer built from commercial off-the-shelf boards provides 44 teraflops of processing power from a series of 4 teraflops boards. The set of images, or training data sets, are specific to an application, and along with the weightings of the network they represent a significant investment in intellectual property for AI developers. One data set might concentrate on people, while another might focus on the types of vehicles on a road, learning the difference between vans, family cars and sportscars, for example. Some developers are even using deep learning for navigation, providing images of a city with landmark recognition instead of a map system. The images can come from actual still pictures or video frames that have been pre-processed to tag the image content. Some types of neural networks are more suited to still images, while others that have more memory in the computation nodes are suited to video streams. AI systems | Focus Unmanned Systems Technology | April/May 2017 The Poplar development system allows the activity of artificial intelligence to be visualised to show the performance of different areas of the network in a way that is similar to a scan of the human brain (Courtesy of Graphcore)