In-memory computing meets Neuromorphic Event-Driven Vision

The intersection of in-memory computing (IMC) and neuromorphic I event-driven vision represents a significant frontier in the Internet of Video Things, as highlighted in the collaborative efforts at the Brain Lab @ CityU HK. This presentation highlights how two aspects of neuromorphic systems--event-driven perception and in-memory computing can be combined to develop very efficient visual perception systems. In particular, event-based image sensors enable data reduction at the source by being sensitive to change in temporal contrast. In this work, the image buffer used to store these data is used to also process it. It is shown that using SRAM or combinations of SRAM and DRAM as the image buffer, basic operations such as noise removal and region proposal may be accomplished within the memory itself. Hence, the following recognition system may be turned off in blank frames and also focus only on salient regions, thus saving significant energy at the system level. Overall, the convergence of neuromorphic event-driven vision and in-memory computing heralds a paradigm shift in low-power, high-performance computing, particularly in applications such as surveillance and intelligent transportation systems of smart cities.