Memristive Fully Convolutional Network: An Accurate Hardware Image-Segmentor in Deep Learning

As well known, fully convolutional network (FCN) becomes the state of the art for semantic segmentation in deep learning. Currently, new hardware designs for deep learning have focused on improving the speed and parallelism of processing units. This motivates memristive solutions, in which the memory units (i.e., memristors) have computing capabilities. However, designing a memristive deep learning network is challenging, since memristors work very differently from the traditional CMOS hardware. This paper proposes a complete solution to implement memristive FCN (MFCN). Voltage selectors are firstly utilized to realize max-pooling layers with the detailed MFCN deconvolution hardware circuit by the massively parallel structure, which is effective since the deconvolution kernel and the input feature are similar in size. Then, deconvolution calculation is realized by converting the image into a column matrix and converting the deconvolution kernel into a sparse matrix. Meanwhile, the convolution realization in MFCN is also studied with the traditional sliding window method rather than the large matrix theory to overcome the shortcoming of low efficiency. Moreover, the conductance values of memristors are predetermined in Tensorflow with ex-situ training method. In other words, we train MFCN in software, then download the trained parameters to the simulink system by writing memristor. The effectiveness of the designed MFCN scheme is verified with improved accuracy over some existing machine learning methods. The proposed scheme is also adapt to LFW dataset with three-classification tasks. However, the MFCN training is time consuming as the computational burden is heavy with thousands of weight parameters with just six layers. In future, it is necessary to sparsify the weight parameters and layers of the MFCN network to speed up computing