Neuromorphic computing, a VLSI realization of neuro-biological architecture, is inspired by the working mechanism of human-brain. As an example of a promising design methodology, synapse design can be greatly simplified by leveraging the similarity between the biological synaptic weight of a synapse and the programmable resistance (memristance) of a memristor. However, programming the memristors to the target values can be very challenging due to the impact of device variations and the limitation of the peripheral CMOS circuitry. A quantization process is used to map analog weights to discrete resistance states of the memristors, which introduces a quantization loss. In this thesis, we propose a regularized learning method that is able to take into account the deviation of the memristor-mapped synaptic weights from the target values determined during the training process. Experimental results obtained when utilizing the MNIST data set show that compared to the conventional learning method which considers the learning and mapping processes separately, our learning method can substantially improve the computation accuracy of the mapped two-layer multilayer perceptron (and LeNet-5) on multi-level memristor crossbars by 4.30% (11.05%) for binary representation, and by 0.40% (8.06%) for three-level representation.
展开▼
