Neural Network Challenges Sample Clauses

Neural Network Challenges. The last 5 years, with the development of Deep Neural Networks (XXXx), neural networks became the best way to analyse information with a high quality of results and low error rates. Even though many kind of neural networks have been created, in the domain of image and video analysis, the main basic computation bloc is a convolution. These neural networks are therefore called Convolutional Neural Networks (CNNs). The images, or the video-frames, are processed through several layers of the network, each layer being in charge of extracting meaningful details from the images. Figure 5 shows a schematic view of a convolutional neural network of five layers. The neu- rons are schematised with a black circle in the layer. The connections between the layers are schematised with an horizontal arrow. The first layer, the input layer, is in charge of extracting the pixels from the image and may apply a first convolution or a filter to the image. The last layer is the output layer. It is in charge of collecting the information from the neural network. The internal layers between the input and output layers are in charge of splitting and partly characterising the image through a list of features and then combining the features to identify and classify the information extracted from the image. Each neuron on one layer provides a set of outputs that are sent to other neurons on the next layer. Each of such a connection between two neurons from two layers is weighted. The set of weights of the connections between the neurons and the weights of the convolutions are the parameters of the neural network. While the performance of these networks is getting higher at each generation, it is difficult to make them run in real-time on processors targeted for embedded products [46]. There are two main constraints to achieve this goal: (1) the amount of computation to process the convolutions is too high to be performed at real-time and (2) the available memory on embedded processing platforms is too small to fit both the data and the parameters of the network. Neural networks are, however, very static. Once a network is designed and parametrised "A smiling Face" Input Image Output Information Input Layers Hidden Layers Output Layers Figure 5: Basic structure of a Convolutional Neural Network - CNN through a learning phase, its implementation and the data path between the layers is known and predictable. This characteristic makes it possible for hardware designers to finetune ar- ch...