Kohonen self organizing maps
- 1. KOHONEN SELF ORGANIZING MAPS
- 2. History of kohonen som Developed in 1982 by Tuevo Kohonen, a professor emeritus of the Academy of Finland Professor Kohonen worked on auto-associative memory during the 70s and 80s and in 1982 he presented his self-organizing map algorithm
- 3. History of Kohonen SOMs His idea on Kohonen SOM only became famous much later in 1988 when he presented a paper on “The Neural Phonetic Typewriter” on IEEE computer that his work became widely known Since then many excellent papers and books have been made on SOM
- 4. What are self organizing maps? •Are aptly named “Self-Organizing” because no supervision is required. •SOMs learn on their own through unsupervised competitive learning. •They attempt to map their weights to conform to the given input data.
- 5. What are self organizing maps? •Thus SOMs are neural networks that employ unsupervised learning methods, mapping their weights to conform to the given input data with a goal of representing multidimensional data in an easier and understandable form for the human eye. (pragmatic value of representing complex data)
- 6. What are self organizing maps? •Training a SOM requires no target vector. A SOM learns to classify the training data without any external supervision.
- 7. The Architecture •Made up of an input nodes and computational nodes. •Each computational node is connected to each input node to form a lattice.
- 8. The Architecture •There are no interconnections among the computational nodes. •The number of input nodes is determined by the dimensions of the input vector.
- 9. Representing Data •Weight vectors are of the same dimension as the input vectors. If the training data consists of vectors, V, of n dimensions: V1,V2,V3...Vn •Then each node will contain a corresponding weight vector W, of n dimensions: W1,W2,W3...Wn
- 10. A sample SOM
- 11. Terms used in SOMs •vector quantization -This is a data compression technique. SOMs provide a way of representing multidimensional data in a much lower dimensional space; typically one or two dimensions
- 12. Terms used in SOMs… •Neighbourhood •Output space •Input space
- 13. EXPLANATION: How Kohonen SOMs work The SOM Algorithm •The Self-Organizing Map algorithm can be broken up into 6 steps •1). Each node's weights are initialized. •2). A vector is chosen at random from the set of training data and presented to the network.
- 14. EXPLANATION: The SOM Algorithm… 3). Every node in the network is examined to calculate which ones' weights are most like the input vector.The winning node is commonly known as the Best Matching Unit (BMU).
- 15. EXPLANATION: The SOM Algorithm… •4).The radius of the neighbourhood of the BMU is calculated.This value starts large.Typically it is set to be the radius of the network, diminishing each time- step.
- 16. EXPLANATION: The SOM Algorithm… •5). Any nodes found within the radius of the BMU, calculated in 4), are adjusted to make them more like the input vector (Equation 3a, 3b).The closer a node is to the BMU, the more its' weights are altered •6). Repeat 2) for N iterations.
- 17. Computation of scores •The Function for calculating the score inclusion for an output node is known as : 𝑖 𝑛𝑖 − 𝑤𝑖𝑗 2 •Thus to calculate the score for inclusion with output node i:
- 18. Computation of scores… •To calculate the score for inclusion with output node j: (0.4 − 0.3)2 + (0.7 − 0.6)2 = 0.141
- 19. The Winning Node •Node j becomes the winning node since it has the lowest score. •This implies that its weight vector values are similar to the input values of the presented instance. •i.e.The value of node j is closest to the input vector. •As a result, the weight vectors associated with the winning node are adjusted so as to reward the node for winning the instance.
- 20. Concluding the tests • Both of these are decreased linearly over the span of several iterations and terminates after instance classifications do not vary from one iteration to the next • Finally the clusters formed by the training or test data are analysed in order to determine what has been discovered
- 21. NEIGHBORHOOD ADJUSMENTS • After adjusting the weights of the winning node, the neighbourhood nodes also have their weights adjusted using the same formula • A neighbourhood is typified by a square grid with the centre of the grid containing the winning node. • The size of the neighbourhood as well as the learning rate r is specified when training begins
- 22. ILLUSTRATION: A Color Classifier •Problem: Group and represent the primary colors and their corresponding shades on a two dimensional plane.
- 23. A Color classifier: Sample Data •The colors are represented in their RGB values to form 3-dimensional vectors.
- 24. A Color classifier: Node Weighting •Each node is characterized by: •Data of the same dimensions as the sample vectors •An X,Y position
- 25. A Color Classifier: The algorithm Initialize Map Radius = d Learning rate = r For 1 to N iterations Randomly select a sample Get best matching unit Scale neighbors Adjust d, r appropriately End for
- 26. A Color classifier:The Layout
- 27. A Color classifier: Getting a winner •Go through all the weight vectors and calculate the Euclidean distance from each weight to the chosen sample vector •Assume the RGB values are represented by the values 0 – 6 depending on their intensity. •i.e. Red = (6, 0, 0); Green = (0, 6, 0); Blue = (0, 0, 6)
- 28. A Color classifier: Getting a winner… •If we have colour green as the sample input instance, a probable node representing the colour light green (3,6,3) will be closer to green than red. • Light green = Sqrt((3-0)^2+(6-6)^2+(3-0)^2) = 4.24 Red = Sqrt((6-0)^2+(0-6)^2+(0-0)^2) = 8.49
- 29. A COLOR CLASSIFIER: DETERMINING THE NEIGHBORHOOD • Since a node has an X – Y position, it’s neighbors can be easily determined based on their radial distance from the BMU coordinates.
- 30. A COLOR CLASSIFIER: DETERMINING THE NEIGHBORHOOD… • The area of the neighbourhood shrinks over time with each iteration.
- 31. A Color classifier: Learning •Every node within the BMU's neighbourhood (including the BMU) has its weight vector adjusted according to a pre-determined equation •The learning rate is decayed over time.
- 32. A Color classifier: Learning •The effect of learning should be proportional to the distance a node is from the BMU. •A Gaussian function can be used to achieve this, whereby the closest neighbors are adjusted the most to be more like the input vector.
- 33. A Color classifier: Sample output • Classification on a 40 x 40 SOM
- 34. Current Applications • WEBSOM: Organization of a Massive Document Collection
- 35. Current Applications… •Classifying World Poverty