Classification using Dopant Network Processing Units

Artificial intelligence (AI) is becoming increasingly important in everyday life. It has led to applications such as face recognition and virtual personal assistants in the form of Siri, Alexa and Cortana. AI is even capable of outperforming humans at complex games such as Go and Dota2. These applications are implemented using artificial neural networks (ANNs). However, the Von Neumann bottleneck limits the speed and size of ANN due to the communication between the memory and processor. In this thesis, we exploit the complex behaviour of disordered networks with a complex non-linear response to implement ANN benchmark tasks in novel hardware.

First, we show that these networks are compatible with the evolution-in-materio (EIM) approach. EIM uses a so-called genetic algorithm (GA) to train complex material systems towards the desired input-output response. The disordered networks used here are dopant network processing units (DNPUs), networks of dopants in a semiconductor that, when operated at the correct temperature and dopant concentration, can have a highly tuneable non-linear response to an externally applied electric field. The DNPUs in this work are made from boron/arsenic doped silicon. Furthermore, we show that DNPUs can be interconnected to perform more complex tasks and that, by utilising the RC framework with DNPUs, it is possible to extract more complex nonlinear behaviour from a single DNPU. Finally, we show that the global tuneability combined with a flexible geometry of DNPUs allows us to reduce the number of stored parameters, and therefore less communication between the memory and processor, needed to perform nonlinear ANN tasks as compared to networks implanted in software. Using this material platform as an example, we emphasize the importance of researching novel material systems to develop intelligent matter capable of both learning complex behaviour and incorporating the memory required to store the learned parameters.