Ambient assisted living deep learning development platforms

Description

Assistance and surveillance can be needed at home to provide guidance and help to elderly people. This can become an uninterrupted activity when a disease, such as dementia, Parkinson or Alzheimer, appears in their lives. Indeed, common daily tasks, taking a bath, cooking or even eating can suddenly develop a dangerous situation. In this context, non-invasive and automated techniques are essential to guarantee at the same time the free will. Many techniques are based on recognition mechanisms, mathematical or algorithmic, such as artificial neural networks (ANN). In particular, the Convolutional Neural Networks (CNN) provide the best solutions to non-linear and complex problems needing massive operations during image or natural language recognition. The challenges are related to the generalization and the over-learning associated to the learning and prediction steps. CNNs incorporate constraints and achieve a degree of invariance shift and deformation using three ideas: local receptive areas, shared weights and sub-sampling space. The use of shared weight reduces the number of parameters in the system facilitating generalization. Thus, CNN has been applied in several applications, we can mention: number recognition, object recognition, voice recognition, machine translation, video surveillance, mobile robotics vision, or search engines pictures. In this context, Deep Learning neural networks, characterized by a large number of hidden layers, allowed to overcome the over-learning problem. They build automatically a representation of increasing high-level, layer by layer, and subset of neurons connected between layers. Great progress has been made recently in the development of high performance (HPC) systems for integration and start-up of the CNN networks based on multi-core technology. The objective of this proposal is to conceive a development environment for automatic generation of CNNs including executable model, training, validation and prediction using the most convenient architecture in terms of execution speed, power processing, memory requirements and power consumption. The tasks to consider are: study of Ambient assisted Living requirements and state-of-art recognition mechanisms, study and propose a CNN based solution, study regularity, scalability and generalization aspects of the proposed CNN, evaluate performance requirements considering multicore, GPU and FPGAs as simulation, training and target architectures.