Neural Networks Dr. Peter Phillips The Human Brain (Recap of week 1) A Classic Artificial Neuron (Recap cont.) X1 W1 X2 W2 f(Sj) Output W3 X3 Sj Unsupervised Learning • Today’s lecture will consider the use of Self Organising Map (SOM) and Unsupervised Learning • Recall that Supervised Learning matches inputs to outputs. • Unsupervised Learning Classifies the data into classes The Biological Basis for Unsupervised Neural Networks • Major sensory and motor systems are ‘topographically mapped’ in the brain –Vision: retinotopic map –Hearing: tonotopic map –Touch: somatotopic map Kohonen Self-Organising Maps • The most famous unsupervised learning network is the Kohonen Network. • Neural network algorithm using unsupervised competitive learning • Primarily used for organization and visualization of complex data Teuvo Kohonen Understanding the Data Set • A good understanding of the data set is essential to use a SOM – or any network for that matter • A ‘distance measure’ and/or suitable rescaling must be defined to allow meaningful comparison • The data must be of good quality and must be representative of the application area SOM - Architecture j 2d array of neurons wj1 wj2 wj3 x1 x2 x3 wjn Weighted synapses ... xn Set of input signals (connected to all neurons in lattice) Finding a Winner (2) • Euclidean distance between two vectors a and b, a = (a1,a2,…,an), b = (b1,b2,…bn), is calculated as: d a, b a bi 2 i i • i.e. Pythagoras’ Theorem • Other distance measures could be used, e.g. Manhattan distance Euclidean distance SOM Parameters • The learning rate and neighbourhood function define to what extent the weights of each node are adjusted Neighbourhood function Degree of neighbourhood Degree of neighbourhood 1 1 Time 0.5 0.5 8 Distance from winner Time 10 6 4 2 -2 -4 -6 -8 -1 0 10 8 6 4 2 0 -2 -4 -8 -1 0 -6 Distance from winner 0 0 0 Data For Tutorial Work • Data collected from a UHT plant at Leatherhead • Consists of 300 cases use 150 for Training and 150 for Testing • Data collected with plant running in normal state, during cleaning of exchangers and with fault Tutorial 2 (UHT Plant Data) UHT plant cleaned and after a period of norm al operation returning to fault condition 1.2 1 0.8 0.6 0.4 0.2 0 1 15 29 43 57 71 85 99 113 127 141 155 169 183 197 211 225 239 253 267 281 295 309 323 337 351 365 379 393 -0.2 Cleaning Normal Fault My settings First run 70 epochs – learning rate 0.6 to 0.1 Second run 50 epochs – learning rate constant 0.1 First run Neighbourhood kept to 1 Second run Neighbourhood start 1 end 0 Trajan Classification SOM – New Data • A trained SOM can be used to classify new input data • The input data is classified to the node with the ‘best’ or ‘closest’ weights • Previous knowledge of other data samples assigned to the same class enable inferences to be made about the ‘new’ input sample