The signal and image processing, as well as classic AI methods, were researched at EDI long before the recent AI hype. In the reporting period we have worked on such image processing tasks as medical image analysis to segment muscle fibers in images, classify skin lesions into malignant and benign, identify people by palmprint and palm vein images, detecting vehicles and other objects in Intelligent Transportation Systems, diameter control for silicon rod growth, etc.
In 2015 we also began research in the Deep Learning field. EDI built a modern HPC server for its time with 12 NVIDIA Tesla k40 and 4 Tesla k20 accelerators. The HPC will soon be upgraded with newer and more powerful GPUs. In the beginning, we researched and applied supervised deep learning using popular convolutional and recurrent neural networks on such tasks as video segmentation for self-driving cars, accurate number plate recognition, detection of crops and weed on the field, etc.
From the research point of view more interesting are the directions that our AI field is branching into. One of them is efficient AI, in which we are embedding the AI inference capabilities on computationally limited devices, such as object detection in videos on a Raspberry Pi computer. For the cases, when training data is not available or is very limited, we are developing data generation methods. Our methods range from semi-automatic labeling methods, using 2D images of the objects to compile training examples, using 3D models and physics engines to render annotated images, and applying generative adversarial networks. Also, explainable, trustworthy, and fail-safe AI is our current interest – we began this direction while working on the perception system of our self-driving car, and currently, we are aiming to use this approach on medical tasks. Finally, in order to glimpse the possible next big thing in AI, we are researching more brain-like learning systems than current simplified artificial neural networks.
Areas: Mobility, Production, Health, Space
EDI expertise in AI (shortly):
• Automatic data generation/labelling methods
• Machine Learning (ML) / Deep Learning (DL)
• Supervised (Convolutional neural network (CNN), Recurrent neural network (RNN), LSTM)
• Unsupervised (e.g. Variational)
• Generative adversarial networks (GAN)
• Transparent/explainable AI
• Embedded intelligence (AI implementation in embedded systems (FPGA, SoC, GPU, etc.))
AI example applications:
• AI based fail-safe perception system for self-driving cars
• AI based traffic monitoring
• AI based robotic arm for pick-up and classification of randomly dropped objects
• AI based human motion analysis
• AI-based classification of various plants and weeds
• etc.
We have our own High-Performance Computing servers for NN training!