Key points are not available for this paper at this time.
Convolutional neural networks (CNN) achieve state-of-the-art results in the field of visual perception, drastically changing the traditional computer-vision framework. However, the movement of massive amounts of data prevents CNN’s from being integrated into low-power IoT devices. The recently proposed binaryweight network (BWN) reduces the complexity of computation and amount of memory access. A conventional digital implementation, which is composed of separate feature/weight memories and a multiply-and-accumulate (MAC) unit, requires large amounts of data to be moved 3. To reduce power the weight memory and the computations are integrated together, into an in-memory computation architecture 1, 2, 5. However, feature data is still stored externally, so data movement has only been partially addressed, especially for BWN. This paper blends feature and partial-weight memory with a computing circuit together, like a sandwich, that achieves significantly less data access (Fig. 24.4.1). It also uses a reconfigurable analog-computation engine, based on pulse-width modulation, that is small and flexible enough to be inserted into the memory.
Yang et al. (Fri,) studied this question.