Performance and Scalability of GPU-Based Convolutional Neural Networks

A committee of neural networks for traffic sign classification

“…Most previous GPUbased CNN implementations [11], [12] were hard-coded to satisfy GPU hardware constraints, whereas ours is flexible and fully online (i.e., weight updates after each image). Other flexible implementations [13] are not fully exploiting the latest GPUs. It allows for training large CNNs within days instead of months, such that we can investigate the influence of various structural parameters by exploring large parameter spaces [14] and performing error analysis on repeated experiments.…”

VPRS-Based Regional Decision Fusion of CNN and MRF Classifications for Very Fine Resolution Remotely Sensed Images

“…Moreover, a nonlinear activation (e.g. sigmoid, hyperbolic tangent, rectified linear units) function is taken outside the convolutional layer to strengthen the non-linearity [44]. Specifically, the major operations performed in the CNN can be summarized as:…”

Gender classification: a convolutional neural network approach

“…The CNN also provides partial resistance and robustness to geometric distortions and transformations, and other 2D shape variations [2]. Hence, the CNN is specifically designed to cope with shortcomings of the traditional feature extractor that is characterized by being static, is designed independently of the trainable classifier, and is not part of training procedure [3]. A final benefit of CNNs is that they are relatively easier to train since they have fewer parameters than fully connected MLP neural networks with the same number of hidden layers.…”

“…The concepts of local receptive field, weight sharing, and spatial subsampling mentioned above are the three principle architectural ideas behind the design of a CNN [2,6]. In weight sharing topology, all neurons in a feature map use the same incoming set of weights (kernel weights), and feature extraction is performed by convolving the image with these kernels [3,11].…”