Figure 1
(a) Simple illustration of employed deep learning neural network (DNN) with the vector matrix multiplication (VMM). (b) Normalized required weight values in software based DNN (S-DNN), and weight values of real synapse device. Contrast to the S-DNN, the synapse device has only positive conductance values which is expressed as weight values. (c) Concept of proposed weight shifter to realize negative weight in the hardware based DNN (H-DNN). The positive conductance values of synapse device can be considered as positively shifted weight values. The difference between required weight values of the S-DNN and positively shifted weight values of the H-DNN (w\(_{shift}\)) can be subtracted during the VMM. (d) Distributions of weight values in the S-DNN and H-DNN. The w\(_{shift}\) can be compensated by the weight shifter during VMM.
