COVER STORY u ANALOG DEVICES
Figure 3: A model of the CIFAR network trained with the CIFAR-10 data set
In our initial example, these are the whiskers or ears of a cat. Visualisation of the feature maps - which can be seen in Figure 4 - is not necessary for the application itself, but it helps in the understanding of the convolution.
convolutional layers and pooling layers, which are both utilised to great effect in the training of neural networks. The convolutional layer uses a mathematical operation called convolution to identify patterns within an array of pixel values. Convolution occurs in hidden layers, as can be seen in Figure 3. This process is repeated multiple times until the desired level of accuracy is achieved. Note that the output value from a convolution operation is always especially high if the two input values to be compared (image and filter, in this case) are similar. This is called a filter matrix, which is also
known as a filter kernel or just a filter. The results are then passed into the pooling layer, which generates a feature map – a representation of the input data that identifies important features. This is considered to be another filter matrix. After training – in the operational state of the network – these feature maps are compared with the input data. Because the feature maps hold object class- specific characteristics that are compared with the input images, the output of the neurons will only trigger if the contents are alike. By combining these two approaches, the
CIFAR network can be used to recognise and classify various objects in an image with high accuracy. CIFAR-10 is one specific dataset commonly
used for training CIFAR neural networks. It consists of 60,000 32×32 colour images broken up into 10 classes that were collected from various sources like web pages, newsgroups, and personal imagery collections. Each class has 6,000 images divided equally between training, testing, and validation sets, making it an ideal set for testing new computer vision architectures and other machine learning models. The main difference between convolutional neural networks and other types of networks is the way in which they process data. Through filtering, the input data are successively examined for their properties. As the number of convolutional layers connected in series increases, so does the level of detail that can be recognised.
12 October 2023 Irish Manufacturing The process starts with simple object
properties, such as edges or points, after the first convolution and goes on to detailed structures, such as corners, circles, rectangles, etc., after the second convolution. After the third convolution, features represent complex patterns that resemble parts of objects in images and that usually are unique to the given object class.
In our initial example, these are the whiskers or ears of a cat. Visualisation of the feature maps - which can be seen in Figure 4 - is not necessary for the application itself, but it helps in the understanding of the convolution. Even small networks such as CIFAR consist of hundreds of neurons in each layer and many layers connected in series. The number of necessary weights and biases grows rapidly with increasing complexity and size of the network. In the CIFAR-10 example pictured in Figure 3, there are already 200,000 parameters that require a determined set of values during the training process. The feature maps can be further processed by pooling layers that reduce the number of parameters that need to be trained while still preserving important information.
As mentioned, after every convolution in a
CNN, pooling, often also referred to in literature as subsampling, often occurs. This serves to reduce the dimensions of the data. If you look at the feature maps in Figure 4, you notice that large regions contain little to no meaningful information. This is because the objects do not make up the entire image, but only a small part of it. The remaining part of the image is not used
in this feature map and is hence not relevant for the classification. In a pooling layer, both the pooling type (maximum or average) and the window matrix size are specified. The window matrix is moved in a stepwise manner across the input data during the pooling process. In maximum pooling, for example, the largest data value in the window is taken. All other values are discarded. In this way, the data is continuously reduced in number, and in the
www.irish-manufacturing.com
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50
