Artificial Neural Networks KONG DA, XUEYU LEI & PAUL MCKAY Neural Networks What is Neural Networks An Artificial Neural Network is a computational simulation of a biological neural network. composed of a large number of highly interconnected processing elements(neurons) an information processing paradigm learn by example What is Neural Networks Inspiration from brain Figure1 Neuron structure Figure 2 ANN structure Figure1: http://www.studyblue.com/notes/note/n/biological-foundations-neuron-communication-/deck/1025438 Figure2: http://www.gdl.cinvestav.mx/~edb/students/evazquez/index.html Warren McCulloch and Walter Pitts modeled a Bernard Widrow and simple neural network Marcian Hoff of Stanford Donald Hebb pointed out that with electrical circuits developed models called neural pathways are strengthened "ADALINE" and The IBM research laboratories ledare used in each time that they "MADALINE." 1943 •The first the first effort to Organization simulate a neural The of Behavior recurrent network •ADALINE Could predict the next network bit 1949 Bayesian network 1950’s •MADALINE TheLearning first neural Vector Quantization 1959 network applied a realterm world Longtoshort memory network 1980 "deep learning" gained traction in thephysical neural network 1982 problem Convolutional neural mid-2000s after a publication Hierarchical temporal memory (HTM) networks were introduced in by Geoffrey Hinton and Ruslan …… 1985 John Hopfield of Caltech created a 1980 paper by Kunihiko Salakhutdinov more useful machines by using Fukushima Companies are working bidirectional lines (Hopfield A Boltzmann machine, a on three types of Network) type Mid 2000s neuro chips - digital, of stochastic recurrent analog, and optical neural network is History of ANNs invented by Geoffrey Hinton and Terry Sejnowski NOW History of ANNs Figure 3 MADALINE structure Figure 4 An example Boltzmann machine Figure3:http://www.drdobbs.com/the-foundation-of-neural-networks-the-ad/184402585 Figure4:http://en.wikipedia.org/wiki/Boltzmann_machine Figure5: http://www.schraudolph.org/teach/NNcourse/lstm.html Figure 6 Neural chip http://www.gizmag.com/neuromorphic-chips/28586/ Figure 5 a maximally simple LSTM network Types of ANNs ANNs Feed-forward Neural Networks Singlelayer Multilayer Recurrent Neural Networks Hopfield network Long short term memory network … Others Kohonen selforganizing network … Perceptron 𝑛 x1 x2 w1 𝑋= 𝑥𝑖 𝑤𝑖 𝑖=1 y w2 𝑌= −1, +1, 𝑥<𝜃 𝑥≥𝜃 Activation function Step function Sigmoid function WikiBooks. Artificial Neural Networks/Activation Functions. 25 August 2014. Example 1 Çelebi, Ömer Cengiz. Neural Networks and Pattern Recognition Using MATLAB. Retrieved 25 August 2014. Example 2 Çelebi, Ömer Cengiz. Neural Networks and Pattern Recognition Using MATLAB. Retrieved 25 August 2014. Input signals Output signals Neural network topology Input layer First hidden layer Second hidden layer Output layer Multilayer Interconnected Feed forward/recurrent Deep learning Training Back propagation Learning rate Batch learning Stochastic gradient descent Evolutionary algorithms Reference: LeCun et al. Efficient BackProp. 1998. Application areas Function approximation http://www.eweb.unex.es/eweb/fisteor/santos/sby.html Classification Screenshot from https://www.youtube.com/watch?v=KuPai0ogiHk Application areas Data processing. http://www.cslu.ogi.edu/tutordemos/nnet_recog/recog.html Control http://www.seit.adfa.edu.au/research/details2.php?page_id =410&topic=Adaptive_Flight_Control Application areas Robotics Screenshots from https://www.youtube.com/watch?v=V2ADU8YWIug