SpinalNet: Deep Neural Network With Gradual Input

Deep neural networks (DNNs) have achieved the state of art performance in numerous fields. However, DNNs need high computation times, and people always expect better a lower computation. Therefore, we study human somatosensory system design network (SpinalNet) to achieve higher accuracy with fewer computations. Hidden layers traditional NNs receive inputs previous layer, apply activation function, then transfer outcomes next layer. In proposed SpinalNet, each layer is split into three splits: 1) input split, 2) intermediate 3) output split. Input receives part inputs. The outputs current number incoming weights becomes significantly than DNNs. SpinalNet can also be used as fully connected or classification DNN supports both learning learning. We observe significant error reductions computational costs most Traditional on VGG-5 provided state-of-the-art (SOTA) QMNIST, Kuzushiji-MNIST, EMNIST (Letters, Digits, Balanced) datasets. ImageNet pre-trained initial SOTA STL-10, Fruits 360, Bird225, Caltech-101 scripts training are available at following link: https://github.com/dipuk0506/SpinalNet Impact Statement—Research deep has gained attention from industries academia due their eye-catching performance. enabled machines perform myriad tasks that once only humans could do. Several researchers recently different types accuracy. recent success biologically inspired convolutional miraculous spinal architecture motivated us develop gradual superior several After first online appearance version this paper, applied new datasets reported promising results. may novel applications upcoming years.