Near Sensors Computation based on Embedded Machine Learning for Electronic Skin

Abstract The electronic skin system is usually made of distributed tactile sensors integrated with an embedded electronic system for tactile data decoding. Meaningful information e.g. texture classification and pattern recognition can be decoded from tactile data by employing machine learning methods. Near sensors computation using embedded machine learning algorithms may enable the electronic skin system to be used in various application domains such as wearable Internet of Things devices, prosthetics, and robotics. However, embedding machine-learning algorithms is constrained by the high computational complexity of Machine Learning methods. This poses relevant challenges on 1) real-time operation and 2) very low (e.g. pJ/op) power/energy consumption due to the limited energy budget available on wearable/portable systems. In this perspective, the paper presents our recent achievements in the implementation of embedded machine learning methods for near sensors tactile data processing. The paper provides an overview about the implementation on dedicated hardware platforms. Finally, efficient techniques for embedded machine learning highlighting the challenges and perspectives are discussed with major emphasis on energy-efficient intelligent electronic skin systems.