A Review on Hyperdimensional Computing

Today, machine learning algorithms play a crucial role in analyzing data from interconnected Internet of Things (IoT) devices. With the rise of the Internet of things also known as IoT, numerous applications use machine learning algorithms to perform cognitive tasks. It has enabled a plethora of applications including personalized medical treatments, accurate face detection, self-driving cars, language detection, image recognition and even more applications. However, these algorithms often pose challenges when it comes to real-time learning on resource-constrained IoT devices. They require a lot of computational resources such as power and memory which makes them difficult to use on today's real-life embedded devices. To address these issues, a new computing paradigm called Hyperdimensional computing (HDC) has emerged. Utilizing HD vectors, known as ‘’hypervectors,” offers a brain-inspired alternative to conventional numerical computing. The basic HDC pipeline consists of encoding, training and inference stages. Nevertheless, the use of existing HDC solutions on today's real-life embedded devices imposes complex constraints on the overall design. In this paper, challenges faced when using HDC on low-power embedded devices are discussed. Contributions in this area are surveyed on encoding, training and inference methods along with benchmarking metrics and applications in HD classification. The surveyed work provides a comprehensive overview of existing real-time HDC architectures and gives insight into future promising research points in this field.