Supervisors: Dr. Radu Timofte
The exponential growth of social media platforms led to a proportional increase of interest in the field of computer vision, in general, and image classification in particular. On a similar note, the trend in industry towards automation requires the development of computer vision algorithms with high accuracy levels. Apart from the applications that use cloud services, there is an interest in image classification systems that work in an edge computing environment. In such a setting, accuracy is not the only metric that defines the performance of a system. One has to consider the latency of the system and its memory requirements. Taking these constraints into account, we developed Multi-Scale Research-Aware Neural Architecture Search. Our aim was to automate the design of an efficient deep neural network, capable of offering fast and accurate prediction and that could be deployed on a low-memory, low-power system-on-chip. We were therefore able to develop an algorithm that can automate the design of efficient neural networks which can be conveniently integrated in microcontrollers or smartphones, thus successfully handling modern world's necessities.