Artificial intelligence (AI) algorithms are currently one of the most powerful tools for solving the most challenging problems, becoming one of the leading research topics and one of the areas with the largest investment in recent years. Disruptive in nature, AI technologies have brought revolutionary changes in many areas. However, one of the main issues preventing the wider application of AI technologies is the high complexity of AI algorithms, which require very expensive and specialized hardware (servers, clouds). In this classical AI concept with complex algorithms of Artificial Neural Networks (ANN) running on clouds, data has to be sent over the Internet to the cloud, causing latency (preventing real-time AI-based applications) as well as issues with security and privacy. Therefore, one of the main directions of cutting-edge research in the field of AI is bringing AI to the Edge, which means designing AI algorithms of reduced complexity that will enable their deployment on edge devices with constrained resources in terms of memory, processing power, energy consumption and latency. This will allow AI algorithms to be executed on local hardware devices as close as possible to the physical system, thus providing a number of new AI-based applications that could fundamentally change various aspects of human life.
Numerous Edge applications, such as object detection and video monitoring, require constant monitoring of sparse real-world data streams with a tiny power budget. Current applications based on ANNs are very inefficient because the sparsity of the input data is not exploited and the requirements of ultra-low-power edge-based computing are not met. Compared to ANNs, brain-inspired Spiking Neural Networks (SNNs) promise more energy-efficient computation and communication as they communicate through events, so-called spikes, that are rare in time. Beyond the sparsity of network activity, the connectivity between layers of neurons can itself be sparse, which further reduces the memory footprint required to store the connectivity matrix by means of compression techniques and thus impacting silicon area and system cost, as well as the energy cost of moving data. Overall, in the research part, this project seeks opportunities to combine the best of both ANN and SNN worlds by co-designing the models towards an adaptive AI-based acquisition and processing pipeline.
In brief, AIDA4Edge is conceived in a way to exploit the synergy of the expertise of the advanced partners to provide the Faculty of Electronic Engineering, University of Niš (FEEUNI) team with necessary knowledge through comprehensive research activities. In addition, it is conceived to increase performance of FEEUNI in a pervasive way (to enhance networking activities, raise reputation, research profile and attractiveness, to strengthen research management capacities and administrative skills), enabling FEEUNI to become a center of excellence in the field of Edge AI.
Objectives
There are five specific objectives of the AIDA4Edge project
OBJECTIVE 1
To establish the base for sustainable research excellence at FEEUNI through: i) the development of the long-term strategy for scientific excellence; ii) identification and connection with stakeholders to be gathered in the Pool of Stakeholders; iii) collection of knowledge acquired in the field of Edge AI in the form of an Online Knowledge Database; iv) the development of early-stage researchers by virtue of joint supervision; v) the enhancement of educational capacities by including the cutting-edge knowledge about Edge AI in curricula.
OBJECTIVE 2
To enhance the network capacity of FEEUNI, enabling it to become a more active and more visible member of European research networks by strengthening the networking between FEEUNI and the top-class leading project partners from Europe, identified stakeholders and researchers from industry and scientific institutions, through organizing numerous networking events and through joint activities in research and supervision of Ph.D. students.
OBJECTIVE 3
To significantly enhance the scientific excellence capacity of FEEUNI in bringing AI to the Edge through: the comprehensive know-how transfer activities and the exchange of the best practice and experiences, covering a wide range of Edge AI topics.
OBJECTIVE 4
To significantly strength research management capacity and administrative skills of FEEUNI staff through: comprehensive trainings of FEEUNI staff, covering all important aspects of research management and administration, and by setting up Research Management and Administrative Unit to support management and administrative issues for future preparation of project proposals and realization, while empowering teams from FEEUNI to apply even more to calls for project funding.
OBJECTIVE 5
To develop an innovative ultra-low-power processing pipeline for video monitoring on Edge AI-based devices by leveraging the synergies between SNNs and ANNs to maintain the accuracy of conventional ANNs and to exploit the input data sparsity through the asynchronous computational capabilities of bio-inspired SNNs.