Convolutional neural networks Bohan Li and Pengxiang Xu Questions: 1. What is complexity of convolutional layer 2. What is the formula of convolution in discrete form? 3. What are the features of CNN? Outline • Overview • History • Algorithms • Applications • Implementations Results • Open issues • References • Discussion Overview: Neural networks • Nueral networks are computing systems inspired by the biological neural networks that constitute animal brain. Overview: Neural networks • Use BP(Backpropagation) neural networks as example. This algorithm contains two parts, forward propagation and back propagation. (1) Forward propagation Overview: Neural networks

(2) Back propagation: The main idea is to build an function of the error in the forward propagation and the weight 푤. Then we can adjust the weight 푤 until the error become small enough. (3)Application: Image recognition The input of image information is different between people and machines. In people's eyes, we see pictures, while the machine could only see a lots of numbers. Overview: Neural networks Overview: Neural networks • Computer receives the digital array that formed according to the resolution of the images and size. • For example, the pixel of an image is 32 x32, but actually a data set of three 32x32 matrices are needed, 32x32x3 because of RGB color model. Overview: Convolutional neural networks • Generally speaking, such a data set mentioned above is very large for neural network and difficult to calculate. Therefore, people need to use more advanced algorithms to process such data. And then you have the convolutional neural network. • Convolutional neural network (CNN) is a class of deep neural networks, most commonly applied to analyzing visual imagery. History • Lecun proposed the first CNN model called ‘LeNet5’ in 1998. • In 2012, the AlexNet proposed by Krizhevsky makes the convolutional neural network the focus of the academic circle Algorithm: Convolutional Neural Network

What is Convolution?

• It is defined as the integral of the product of the two functions after one is reversed and shifted. 푓 ∗ 푔 푡 = ∫ 푓 푥 푔 푡 − 푥 푑푥

Discrete: ∞ 푓 ∗ 푔 푡 = ෍ 푓 푥 푔 푡 − 푥 푡=− ∞ How about metrix How it works in CNN

Complexity

• M Side length of each convolution kernel output feature map • K side length of each convolution kernel

• 퐶푙−1: The number of channels per convolution kernel, that is, the number of input channels, that is, the number of output channels of the previous layer.

• 퐶푙: The number of convolution kernels that this convolution layer has, that is, the number of output channels. Features • Locally Connected Network • Sharing weights • Pooling Pooling

Eg.Max pooling

Structure of CNN Implementations Results Application • Image detection -Faster R-CNN • Image classification • Image recognition -Facial recognition Open issues • The size and quality of data • Models need to be trained for many times in the future to upgrade and the initial value of weight has a big impact on result. • Usually need GPU to train convolutional neural networks. • Do not have rotation invariance References • van Gerven, Marcel, and Sander Bohte, eds. Artificial neural networks as models of neural information processing. Frontiers Media SA, 2018. • Yann, LeCun. "LeNet-5, convolutional neural networks." (2013). • Fukushima, Kunihiko. "Neocognitron." Scholarpedia 2.1 (2007): 1717. • Fukushima, Kunihiko. "Neocognitron: A self-organizing neural network model for a mechanism of unaffected by shift in position." Biological cybernetics 36.4 (1980): 193-202. • Ongsulee, Pariwat. ", machine learning and ." 2017 15th International Conference on ICT and Knowledge Engineering (ICT&KE). IEEE, 2017. Discuss Question again 1. What are two parts of BP neural network? 2. What is the formula of convolution in discrete form? 3. What are the features of CNN?