2018/02/07 by Md Zahangir Alom, Alom, Md Zahangir, Theodore Josue +9 · 1 citation
Engineering · #Advanced Memory and Neural Computing #Ferroelectric and Negative Capacitance Devices #CCD and CMOS Imaging Sensors
paper · pdf · doi:10.48550/arxiv.1802.02608
In the last decade, special purpose computing systems, such as Neuromorphic\ncomputing, have become very popular in the field of computer vision and machine\nlearning for classification tasks. In 2015, IBM's released the TrueNorth\nNeuromorphic system, kick-starting a new era of Neuromorphic computing.\nAlternatively, Deep Learning approaches such as Deep Convolutional Neural\nNetworks (DCNN) show almost human-level accuracies for detection and\nclassification tasks. IBM's 2016 release of a deep learning framework for\nDCNNs, called Energy Efficient Deep Neuromorphic Networks (Eedn). Eedn shows\npromise for delivering high accuracies across a number of different benchmarks,\nwhile consuming very low power, using IBM's TrueNorth chip. However, there are\nmany things that remained undiscovered using the Eedn framework for\nclassification tasks on a Neuromorphic system. In this paper, we have\nempirically evaluated the performance of different DCNN architectures\nimplemented within the Eedn framework. The goal of this work was discover the\nmost efficient way to implement DCNN models for object classification tasks\nusing the TrueNorth system. We performed our experiments using benchmark data\nsets such as MNIST, COIL 20, and COIL 100. The experimental results show very\npromising classification accuracies with very low power consumption on IBM's\nNS1e Neurosynaptic system. The results show that for datasets with large\nnumbers of classes, wider networks perform better when compared to deep\nnetworks comprised of nearly the same core complexity on IBM's TrueNorth\nsystem.\n