Deep Learning

Deep Learning

Khalid A. Al Afandy, Hicham Omara, Mohamed Lazaar, Mohammed Al Achhab
Copyright: © 2022 |Pages: 41
DOI: 10.4018/978-1-7998-8929-8.ch006
OnDemand:
(Individual Chapters)
Available
$37.50
Current Special Offers
No Current Special Offers
TOTAL SAVINGS: $37.50

Abstract

This chapter provides a comprehensive explanation of deep learning including an introduction to ANNs, improving the deep NNs, CNNs, classic networks, and some technical tricks for image classification using deep learning. ANNs, mathematical models for one node ANN, and multi-layers/multi-nodes ANNs are explained followed by the ANNs training algorithm followed by the loss function, the cost function, the activation function with its derivatives, and the back-propagation algorithm. This chapter also outlines the most common training problems with the most common solutions and ANNs improvements. CNNs are explained in this chapter with the convolution filters, pooling filters, stride, padding, and the CNNs mathematical models. This chapter explains the four most commonly used classic networks and ends with some technical tricks that can be used in CNNs model training.
Chapter Preview
Top

The Anns

The ANNs approach is an algorithm that tries to mimic the human brain in which the use of the ANNs as a supervised classifier is based on the structure of the biological NNs (Shanmuganathan, 2016; Gaur et al., 2021b). The structure of ANNs depends on the data that flow through this network. The accuracy and the performance of the ANNs are highly dependent on the network structure. The ANNs computational rate is high but the network takes a huge time for training and there is some difficulty to choose the network structure because the network design is based on intuition (Du et al., 2016; AlAfandy et al., 2019).

NNs are deemed as nonlinear statistical data modeling tools where the complex relationships between inputs and outputs are modeled. NNs are formed of a sequence of layers; each layer contains a set of neurons. The input and the output layers are the first and the last layers, where the internal layers are treated as the hidden layers. Neurons in the preceding and the succeeding layers are connected by weighted connections called the weights (Du et al., 2016). In the ANNs models, the given dataset inputs and outputs must have a mathematical relation in which the outputs are predicted within discrete function for classification models or continuous function for regression models; this mathematical model is based on the ANNs structure and design. This mathematical model is the calculation of the prediction 978-1-7998-8929-8.ch006.m01 which is a function of weights 978-1-7998-8929-8.ch006.m02, bias 978-1-7998-8929-8.ch006.m03, and input features 978-1-7998-8929-8.ch006.m04 (Bengio et al., 2017).

Figure 1.

One node ANN structure

978-1-7998-8929-8.ch006.f01

As shown in figure 1, this example is an ANN structure which is consists of one node with input features 978-1-7998-8929-8.ch006.m05, the output in this model is calculated as (Bengio et al., 2017):

978-1-7998-8929-8.ch006.m06
(1)
978-1-7998-8929-8.ch006.m07
(2)
978-1-7998-8929-8.ch006.m08
(3) where 978-1-7998-8929-8.ch006.m09 is the ANN output, 978-1-7998-8929-8.ch006.m10 is the dataset input, 978-1-7998-8929-8.ch006.m11 is the activation function, 978-1-7998-8929-8.ch006.m12 is the bias, and 978-1-7998-8929-8.ch006.m13 is the weight matrix.

Complete Chapter List

Search this Book:
Reset