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

A convolutional neural network, also called a feed-forward, shift-invariant or space- invariant artificial neural network, is most commonly used for image analysis applications, as they have shown the best results. For applications such as speed recognition meanwhile, a recurrent neural net has internal feedback loops that allow it to use internal memory to process arbitrary sequences of inputs. Another type, particularly useful for video from cameras, is called long short-term memory, which is a type of recurrent neural network that adds memory so it can ‘remember’ changes over time. Then there is a recursive neural network, which uses the same matrix of weightings across the whole network rather than individual weightings for each computational node, so the whole matrix changes during training. These are being used increasingly for natural language processing and detecting subtle changes within a class of objects. For other areas of a vehicle system design that do not need image analysis, such as path planning, radial basis function and dynamic wave expansion neural nets have been found to be the most effective. Types of neural network 36 neural network. The software can then be fully tested within the system, then downloaded to the vehicle to improve overall system performance. That also means all the vehicles will be running the same version of software, rather than lots of different versions. Research is also being carried out into unsupervised learning that aims to update the nets in the vehicle, but that brings with it the same safety and certification challenges. Implementation How these systems are implemented is a key issue for developers. While there are advances in hardware, the software development process is very different from that in traditional projects. As a result, a new generation of software development kits for the different hardware architectures and neural net software libraries – to support both the training and deployment stages of projects – is emerging. In the past two years the x86 chip architecture has been tweaked so that it can handle the requirements of high-performance neural networks for autonomous systems, specifically in handling multiply accumulate (MAC) instructions in a single cycle. These server-class processors, with eight or 16 cores, now operate with a maximum power dissipation of 40 W, the maximum for driverless car and even UAV designs. They are being used on boards using the COM Express Type 10 format for control systems, and on VPX boards for AI image analysis in larger UAVs. Their advantage is that they provide a well-known development environment, high-speed memory and comms interfaces, and the long lifetime support needed for automotive designs. The dominant technology currently for deep learning is the GPU. This is optimised for MAC operations, and devices have up to 256 highly April/May 2017 | Unmanned Systems Technology Focus | AI systems Above left: The AlexNet deep neural network architecture was introduced in 2012 and marked a breakthrough architecture for deep neural networks with much more accuracy than other methods (Courtesy of Graphcore) Above right: In 2015, Microsoft Research introduced the ResNet architecture for building deep neural networks. The clearly visible clustering is the result of intensive communication between processes in each layer of the network, with lighter communication between layers (Courtesy of Graphcore)