Abstract
Deep neural networks offer advanced procedures for many learning tasks because of the ability to extract preferable features at every network layer. The evolved efficiency of extra layers inside a deep network will come at the expense of appended latency and power consumption in feedforward inference. As networks continue to grow and deepen, these outcomes become exceedingly prohibitive for energy-sensitive and real-time software. To overcome this problem, we propose the Side Road Network (SRN), an innovative deep network structure that is enhanced with further side road (SR) classifiers. The SR classifiers are trained by Pseudoinverse learning algorithm (PIL). The PIL algorithm does not integrate crucial user-dependent parameters such as momentum constant or learning rate. The SRN structure allows the prediction of results for a major portion of test samples to exit the network earlier via these SR classifiers since samples can be inferred with certainty. We analyze SRN structure using different models such as VGG, ResNet, WRN, and MobileNet. We evaluate the performance of SRN on three image datasets-CIFAR10, CIFAR100, and Tiny ImageNet-and show that it can improve the model prediction at earlier layers.