Neural Networks

Fall 2003

Proposal | Introduction | Basics | Algorithm | Optimizations | Conclusions | Applications | Demonstration | References |

For my term project I will research and implement a Self-organizing Map (SOM). I will submit an introductory guide to SOMs with a brief critique on its strengths and weaknesses. In addition, I will write a program that implements and demonstrates the SOM algorithm in action. This program will be for tutorial purposes and will simply show how a SOM maps 3-dimensional input down to a 2-dimensional grid where geometric relationships between points indicate similarity.

Back to top

A Self-organizing Map is a data visualization technique developed by Professor Teuvo Kohonen in the early 1980's. SOMs map multidimensional data onto lower dimensional subspaces where geometric relationships between points indicate their similarity. The reduction in dimensionality that SOMs provide allows people to visualize and interpret what would otherwise be, for all intents and purposes, indecipherable data. SOMs generate subspaces with an unsupervised learning neural network trained with a competitive learning algorithm. Neuron weights are adjusted based on their proximity to "winning" neurons (i.e. neurons that most closely resemble a sample input). Training over several iterations of input data sets results in similar neurons grouping together and vice versa. The components of the input data and details on the neural network itself are described in the "Basics" section. The process of training the neural network itself is presented in the "Algorithm" section. Optimizations used in training are discussed in the "Optimizations" section.

SOMs have been applied to several problems. The simple yet powerful algorithm has been able to reduce incredibly complex problems down to easily interpreted data mappings. The main drawback of the SOM is that it requires neuron weights be necessary and sufficient to cluster inputs. When an SOM is provided too little information or too much extraneous information in the weights, the groupings found in the map may not be entirely accurate or informative. This shortcoming, along with some other problems with SOMs are addressed in the "Conclusions" section.

The rest of this site is dedicated to interesting applications of SOMS that I have found. These range from powering search engines and speech pattern recognition software to analyzing world poverty. A short description of these projects and how they incorporated SOMs in them can be found in the "Applications" section.

Lastly, I have coded an OpenGL program designed to demonstrate an SOM for tutorial purposes. The executable, source code, and documentation on what the program does and how to use it can be found in the "Demonstration" section.

The sources I used for this final project are listed in the "References" section.

Back to top

## Input Data

For the sake of clarity, I will describe how an SOM works by using a simple example. Imagine a SOM that is trying to map three dimensional data down to a two dimensional grid. In this example the three dimensions will represent red, blue, and green (RGB) values for a particular color. An input data set for this problem could look something like this:

For this data set, a good mapping would group the red, green, and blue colors far away from one another and place the intermediate colors between their base colors (e.g. yellow close to red and green, teal close to green and blue, etc).

## Neural Network

The neural network itself is a grid of neurons. Each neuron contains a weight vector representing its RGB values and a geometric location in the grid. These weight vectors will be used to determine the "winning" neuron for each input and are updated based on its location during the training process. An initialized neural network for this problem could look something like this:

The squares in this picture represent the neurons and the descending verticle lines show their intialized RGB values. Neuron weights can either be randomly initialized or generated beforehand. In most cases random initialization is sufficient as the SOM will inevitably converge to a final mapping. However, pregenerating neuron weights can drastically improve convergence time and will be discussed in the "Optimizations" section.

That covers the basics of a SOM. The input is a set of multi-dimensional vectors and the neural network is a grid of neurons that each contain a weight vector of the same dimensionality as the input vector. Now to explain the algorithm used to train the neural network.

Back to top

The SOM learning algorithm is relatively straightforward. It consists of initializing the weights as mentioned above, iterating over the input data, finding the "winning" neuron for each input, and adjusting weights based on the location of that "winning" neuron. A pseudocode implementation is provided below:

Initialize weights

For 0 to X number of training epochs

Select a sample from the input data set

Find the "winning" neuron for the sample input

Adjust the weights of nearby neurons

End for loop## Finding the "Winning" Neuron

The algorithm used to find the "winning" neuron is a Euclidean distance calculation. That is, the "winning" neuron

nof dimensiondfor sample inputv(also of dimensiond) would be the one which minimized the equation:

So for the three dimensional RGB example the "winning" neuron would be the one which minimized:

## Finding / Adjusting Neighbors

Adjusting the weights of nearby neurons can be done in several ways. The algorithm for doing so involves first determining which neurons to scale then how much to scale them by. The neurons close to the "winning" neuron are called its "neighbors". Determining a neuron's "neighbors" can be achieved with concentric squares, hexagons, and other polygonal shapes as well as Gaussian functions, Mexican Hat functions, etc. Generally, the neighborhood function is designed to have a global maxima at the "winning" neuron and decrease as it gets further away from it. This makes it so that neurons close to the "winning" neuron get scaled towards the sample input the most while neurons far away get scaled the least (or in some cases in the mexican hat function, scaled away from the sample input), which creates groupings of similar neurons in the final map. Below is an image of a three dimensional Gaussian function and a Mexican Hat function:

The neighborhood function's radius is often decremented over time so that the amount of "neighbors" decreases as training progresses. This is done to help neurons initially adjust their weights to roughly where they want to be then allow them to converge without being dramatically influenced by "winning" neurons that are far away.

The amount to adjust each "neighbor" by is determined by the following formula:

Similar to the neighborhood function, the amount to adjust each "neighbor" by can also be adjusted over time so that initial neuron weight adjustment is drastic but is less easily influenced as training progresses. This is done for the same reasons mentioned for adjusting the neighborhood function's radius.

That covers the SOM algorithm for training its neural network. Neurons converge to final weight values through a competitive learning scheme that adjusts them to resemble nearby "winning" neurons. This process generates groups of similar neurons in the final map

Back to top

Professor Teuvo Kohonen, along with a group of researchers at the Neural Networks Research Center in Helsinki University of Technology, developed a few optimization techniques for SOM training. Their project, WEBSOM, was designed to organize massive document collections in real time using a SOM. In order to accomplish this the research group had to develop methods for speeding up SOM training time. The optimizaitons they came up with are detailed in their report and involve reducing the dimensionality of input data, initializing neuron weights closer to their final state, and restricting the search for "winning" neurons. It is important to note that these optimization techniques were designed specifically for their project and do not guarantee the same clustering accuracy when used in other problems. However, the improvement in training time does carry across all non-trivial SOMs.

## Reducing Dimensionality

To provide some background information into the inspiration for reducing input data dimensionality, I should mention that the initial input data for the WEBSOM project was about 43,000-dimensional; each dimension corresponded to the frequency of occurence of a particular word in a document. Furthermoer, the original WEBSOM neural network contained over 1,000,000 neurons. Because of the neural network's large size and the large dimensionality of the weight vectors, training was a long and computationally heavy task.

Fortunately, the WEBSOM researchers came up with a way to reduce the dimensionality of their data. They were able to do this by utilizing the Random Projection algorithm. Random Projection takes

d-dimensional dataXand projects it to ak-dimensional subspace (wherek<<d) by mulitplying a randomkxdmatrixR, whose columns are normally distributed vectors of unit length, with the input dataX. This produces a projected matrix of dimensionk:

The reason Random Projection works is explained in the Johnson-Lindenstrauss lemma:

The "distance" between two points refers to the dot product of two column vectors. Basically the Johnson-Lindenstrauss lemma says the similarity between a pair of projected vectors will be the same, on average, as the similarity of the corresponding pair of original vectors. In addition, the WEBSOM researchers argued that if the original similarities between the points were already suspect then there would be little reason to preserve their relationships entirely. This justification allowed them to drastically cut down the dimensionality of the input data and neuron weight vectors (from 43,000 to 500) without a noticeably drop in accuracy (from 60.6% to 59.1%)."If points in a vector space are projected onto a randomly selected subspace of suitably high dimension, then the distances between the points are approximately preserved."

## Pregenerated Neuron Weights

Another technique the WEBSOM researchers utilized was a method of initializing neuron weight values closer to their final state. This was accomplished by estimating larger maps based on the asymptotic convergence values of a smaller map. The researchers argued that constructing a smaller neural network, training it with the input data, then interpolating/extrapolating data from the smaller, converged map would generate a blueprint for the construction of the larger map. This allowed the map to start its training closer to the converged state which sped up the training process.

## Restricting the Search for "Winning" Neurons

The last optimization technique mentioned in the WEBSOM paper was a method of restricting the search for "winning" neurons as training progressed. Initially, the search for the "winning" neuron must be thorough in order to find the best fit neuron in the map. However, the WEBSOM researchers argued that once the map starts to become smoothly ordered, the algorithm should be able to start restricting the search for "winning" neurons to neurons close to the old "winner". The researchers felt that after a significant amount of training, the clusters would develop to the point where the location of "winning" neurons would not move around much.

The gradual restriction of the "winning" neuron search turned out to be significantly faster than performing exhaustive searches over the entire map for every input. To ensure matches were a global best, a full search was performed intermittently.

Back to top

## Advantages

The main advantage of using a SOM is that the data is easily interpretted and understood. The reduction of dimensionality and grid clustering makes it easy to observe similarities in the data. SOMs factor in all the data in the input to generate these clusters and can be altered such that certain pieces of data have more/less of an effect on where an input is placed.

SOMs are capable of handling several types of classification problems while providing a useful, interactive, and intelligable summary of the data. For example, a SOM could be used in place of a Bayesian spam filter. Using a weight vector similar to the one used in the WEBSOM project, a SOM should be able to map e-mails onto a grid with clusters representing spam and not spam. Training would be occur whenever the user marked an e-mail as spam or not spam. The user would be presented with a graphical map of e-mail clusters, allowing him/her to graphically view where an e-mail fell compared to other e-mails they had previously marked as spam / not spam.

Finally, as the WEBSOM project showed, SOMs are fully capable of clustering large, complex data sets. With a few optimization techniques, a SOM can be trained in a short amount of time. Furthermore, the algorithm is simple enough to make the training process easy to understand and alter as needed.

## Disadvantages

The major disadvantage of a SOM, which was briefly mentioned in the introduction, is that it requires necessary and sufficient data in order to develop meaningful clusters. The weight vectors must be based on data that can successfully group and distinguish inputs. Lack of data or extraneous data in the weight vectors will add randomness to the groupings. Finding the correct data involves determining which factors are relevant and can be a difficult or even impossible task in several problems. The ability to determine a good data set is a deciding factor factor in determining whether to use a SOM or not.

Another problem with SOMs is that it is often difficult to obtain a perfect mapping where groupings are unique within the map. Instead, anomalies in the map often generate where two similar groupings appear in different areas on the same map. Clusters will often get divided into smaller clusters, creating several areas of similar neurons. This can be prevented by initializing the map well, but that is not an option if the state of the final map is not obvious.

Finally, SOMs require that nearby data points behave similarly. Parity-like problems such as the XOR gate do not have this property and would not converge to a stable mapping in a SOM.

Back to top

Here are links a few applications of SOMs I found along with a picture of their SOM and a brief summary of their project:

## WEBSOM: Organizing a Massive Document Collection

WEBSOM is a SOM that organizes massive document collections. It was developed by Professor Teuvo Kohonen at the Neural Networks Research Center in Helsinki University of Technology in the late 1990's. The project successfully created a SOM that acted as a graphical search engine, classifying over 7,000,000 patent abstracts based on the frequency of occurence of a set of words.

My Powerpoint presentation on Self-organizing maps and WEBSOM is available here.

## Phonetic Typewriter

The Phonetic Typewriter is a SOM that breaks recorded speech down to phonemes. It was developed also by Professor Teuvo Kohonen but in the late 1980's. The documentation on the project can be found here but requires the user to purchase it in order to view it. My understanding of the article is therefore restricted to the information presented in our lecture slides which state that the Phonetic Typewriter was successful at translating speech into phoneme text in Finnish and Japanese. The developers claim that 92% accuracy was achieved within 10 minutes of training. The picture above is the SOM, which they called a phonetic map with the neurons and the phonemes which they learned to respond to.

## Classifying World Poverty

The World Poverty Classifier is a SOM that maps countries based on 39 "quality-of-life factors" including health, nutrition, education, etc. The two pictures above show the SOM itself along with a key of country acronyms and a map of the world where the countries have been colored in by their corresponding locations in the SOM.

Back to top

## Download

My OpenGL program demonstrating the SOM training algorithm is available here.

Note: The program requires the glut32.dll file which is available here. Place it in the c:\WINDOWS\system32 directory. The glut.dll file is available here if you need it as well.The sourcecode is available here.

## Results

The program behaved very well after minor adjustments to constants in the Gaussian neighborhood function. I obtained a wide spectrum of results depending on my input set and how quickly I decreased the Guassian neighborhood function's influence and radius. The program demonstrates an SOM training on a randomly initialized grid of neurons. It is effective at forming distinct clusters of similar neurons but sometimes has trouble developing the blended transitions between these clusters.

## Using the Program

The program runs on windows machines that support OpenGL and GLUT. It presents the user with a randomly initialized map. Training can be toggled on and off by pressing the '

t' key. A console updates the user on whether training is on or off as well as what training epoch the program is on. The map can also be reset by pressing the 'r' key. In preliminary test runs the map converged within 30 seconds to 1 minute.## Implementation Details

The map is a 100 x 100 grid of neurons each containing a weight vector corresponding to its RGB value. The neuron weights are randomly initialized using the computer's clock time as a seed. Training involves randomly selecting an input from one of 48 unique colors, finding a "winning" neuron using Euclidean distance, and adjusting the "neighborhood" weights. I used a Gaussian function whose amplitude and radius decrease as the number of epochs increases to determine how much to adjust the "neighborhood" weights. Because the neuron weights are randomly initialized, the map converges differently each time. Sometimes the map converges quickly and elegantly while othertimes the map has difficulty in reaching stable equilibrium or filling in the gaps between clusters.

## Screenshots

Back to top

Randomly initialized map

Map trained on shades of red, green, and blue

Map trained on shades of red, green, and blue with some yellow, teal, and pink

Map trained on shades of red, green, and blue with more shades of yellow, teal, and pink

Another map trained on many shades of colors

A well converged map with several clusters trained on 48 distinct colors

1. Germano, Tom. Self-Organizing Maps. http://davis.wpi.edu/~matt/courses/soms/2. Honkela, Timo. Description of Kohonen's Self-Organizing Map. http://www.mlab.uiah.fi/~timo/som/thesis-som.html3. Keller, Bob. Self-Organizing Maps. http://www.cs.hmc.edu/courses/2003/fall/cs152/slides/som.pdf4. Kohonen, Teuvu et al. Self Organization of a Massive Document Collection. Helsinki University of Technology, 20005. Unknown author. SOM Toolbox for Matlab. http://www.cis.hut.fi/projects/somtoolbox/6. Zheng, Gaolin. Self-Organizing Map (SOM) and Support Vector Machine (SVM). http://www.msci.memphis.edu/~giri/compbio/f00/GZheng.htm

Here are the links posted on this website in case you missed them:

Back to top

WEBSOM project

My powerpoint presentation on SOMs and WEBSOM

Slides from class on SOMs

Phonetic typewriter report

Classifying World Poverty project

My SOM program

The sourcecode for my SOM program

My powerpoint presentation on this final project