Using biologically inspired algorithms in the physical world
Using biologically inspired algorithms on 'edge devices', such as cameras and mobile phones, is what Svetlana Minakova's PhD research was all about. She conducted research on Convolutional Neural Networks(CNN). Making these algorithms work in different situations is a complicated task. 'Most design tools do not take into account the limitations of physical devices that CNN algorithms have to work on.'
CNN is an algorithm that can learn from visual data, especially 2D and 3D images. These algorithms are for example used in security cameras. 'Since about 2015, researchers have started developing a design flow for CNN.' Minakova explains that it is not easy to build them. 'These models do not take into account limitations of physical objects. For example, a small device like a camera may not have enough memory, or the battery may not be strong enough to run the algorithm. In my thesis, I tried to find ways to solve this.'
Applying algorithms in the physical world
To solve the problem, Minakova started working on a new type of design flow. 'I tried to change the properties of the CNN algorithm I was working with,' she says. Minakova explains that, among other things, the speed at which a device processes the algorithm can differ. Instead of creating a new algorithm, she adapted the characteristics of existing algorithms.
During this process, Minakova realised that adapting algorithms to ‘edge devices’ is not an easy task. 'I discovered that it is really challenging to find common rules for efficient application of CNN algorithms on different platforms, because the platforms are all very different.'
A step towards better functioning devices
Although Minakova found it difficult to apply these algorithms properly, the results of her research have paid off. By creating a new layer in the design flow, researchers can take into account the platforms they work for, and continue to create better CNN algorithms for digital devices.