Synaptic Weights in a Neuromemristive Radioisotope Classifier

A memristor crossbar’s intrinsic device network dynamics can be harnessed to efficiently conduct radionuclide classification tasks by retrofitting the network with a peripheral CMOS-based architecture that has been structurally and functionally optimized for this classification task. However, the hardware implementation of this classification platform is limited by the physical characteristics of the memristor which has a finite number of states. This renders the employment of traditional neural network learning algorithms, where the weights are not limited to defined states, as an excessively complex task. Hence, this paper tests the limitations on weight resolution and its effect in classification precision when implementing a spiking locally competitive learning algorithm. Both linear and nonlinear weight distributions are examined. The algorithm’s local competitiveness is assessed for the specific application of radionuclide identification. The system is tested using spectra data obtained from the United States National Nuclear Data Center as the classification database dictionary. The platform’s accuracy is measured when test signals with 100, 10 and 1 signal-to-noise ratios are assessed. It has been shown that the system is highly effective for classifying radioisotopes with linear weight distribution even with high levels of noise present. A minor classification accuracy improvement was also observed for weight states distributions with a higher density of values in the low conductivity range. Therefore, it is concluded that a memristor-based radionuclide classifier should have at least 4 possible states mapping the algorithm’s synaptic weights.