FAFU机器学习10-2 ANN课件.pptxVIP

Foundations of Machine LearningArtificial Neural Networks2023/11/4Artificial Neural NetworksLesson 10 - 1 Artificial Neural NetworksArtificial Neural Networks: IntroductionSingle Layer Neural NetworksMultiple Layer Neural NetworksSelf-Organizing Map (SOM)Other Neural Networkssklearn.neural_network2023/11/4Artificial Neural NetworksLesson 10 - 2 Artificial Neural Networks: IntroductionThe brain and the neurons （脑与神经元）Neurons are the building blocks of the brainTheir interconnectivity forms the programming that allows us to solve all our everyday tasksThey are able to perform parallel and fault tolerant computationTheoretical models of how the neurons in the brain work and how they learn have been developed from the beginning of Artificial IntelligenceMost of these models are really simple (but yet powerful) and have a slim resemblance to real brain neurons2023/11/4Artificial Neural NetworksLesson 10 - 3 Artificial Neural Networks: IntroductionA neuron model （神经元模型）1943 年， [McCulloch and Pitts, 1943] 抽象出“M-P 神经元模型”, 在这个模型中， 神经元接收到来自n个其他神经元传递过来的输入信号，这些输入信号通过带权重的连接( connection)进行传递，神经元接收到的总输入值将与神经元的阈值进行比较，然后通过“激活函数” (activation function) 处理以产生神经元的输出。2023/11/4Artificial Neural NetworksLesson 10 - 4 Artificial Neural Networks: Introduction理想中的激活函数是阶跃函数，它将输入值映射为输出值“0” 或者“1”，显然“1” 对应于神经元兴奋， “0” 对应于神经元抑制。然而，阶跃函数具有不连续、不光滑等不太好的性质，因此实际常用Sigmoid作为激活函数，它把可能在较大范围内变化的输入值挤压到(0, 1) 输出值范围内，因此有时也称为挤压函数 (squashing function). 2023/11/4Artificial Neural NetworksLesson 10 - 5 Artificial Neural Networks: IntroductionOrganization of neurons/Networks，神经元组织/网络Usually neurons are interconnected forming networks, there are basically two architecturesFeed forward networks (前馈网络), neurons are connected only in one direction Recurrent networks （递归网络，或者循环网络）, outputs can be connected to the inputsFeed forward networks are organized in layers, one connected to the otherSingle layer neural networks (perceptron networks，感知器网络): input layer （输入层）, output layer （输出层）Multiple layer neural network