An Efficient and Effective Network Is All You Need

The demand for efficient deep AI networks is rising due to increasingly complex artificial intelligence and deep learning tasks. This trend has led to high computational expenses. Deep convolutional networks (ConvNets) are built with pointwise convolutional neural networks (CNNs) (1x1) and spatial CNNs. The pointwise CNNs manage the number of channels, while the spatial CNNs extract feature maps. However, the cost complexity of pointwise CNNs is quite high. This paper introduces a new approach to building a ConvNet efficient and effective network, E<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> Net, architecture by replacing pointwise CNN layers with Squeeze-and-Excitation networks (SENets) and substituting the 2D spatial CNN layer with a parameter-efficient residual 1D CNN (RCN) layer for vision tasks. We demonstrate the effectiveness of our proposed model on brain tumor detection and image classification datasets. The extensive experiment shows that our proposed 44 layers with widening factor 2 model achieves 99.68% and 99.45% validation accuracy for 2 and 3-classes brain tumor datasets, respectively. This model consumes 2.1M parameters and 12.5M FLOPS. Moreover, our proposed light network shows state-of-the-art performance on these datasets. Our model outperformed some relevant architectures for CIFAR benchmarks, SVHN, and tiny-ImageNet datasets.