Self Organizing Maps - Kohonen Maps

A Self Organizing Map (SOM) or Kohonen Map is an unsupervised neural network algorithm based on biological neural models from the 1970s. It uses a competitive learning approach and is primarily designed for clustering and dimensionality reduction. SOM effectively maps high-dimensional data to a lower-dimensional grid enabling easier interpretation and visualization of complex datasets.

👁 Image

It consists of two primary layers:

• Input Layer: Represents the features of the data.
• Output Layer: Arranged as a 2D grid of neurons with each neuron representing a cluster in the data.

Working of Self Organizing Maps (SOM)

Here are the step by step explanation of its working:

1. Initialization

The weights of the output neurons are randomly initialized. These weights represent the features of each neuron and will be adjusted during training.

2. Competition

For each input vector, SOM computes the Euclidean distance between the input and the weight vectors of all neurons. The neuron with the smallest distance is the winning neuron.

Formula :

where

• is the distance for neuron
• and is the number of features.

3. Weight Update

The winning neuron’s weights are updated to move closer to the input vector. The weights of neighboring neurons are also adjusted but with smaller changes.

Formula:

where

• is the learning rate
• ​ is the input feature.

4. Learning Rate Decay

The learning rate α\alphaα decreases over time allowing the map to converge to stable values.

Formula:

5. Stopping Condition

The training stops when the maximum number of epochs is reached or when the weights converge.

Implementation of SOM in Python

Now let’s walk through a Python implementation of the SOM algorithm. The code is divided into blocks for clarity.

1. Importing Required Libraries

We will use the math library to compute the Euclidean distance between the input vector and the weight vector.

2. Defining the SOM Class

In this class, we define two important functions, winner() to compute the winning neuron by calculating the Euclidean distance between the input and weight vectors of each cluster and update() to update the weight vectors of the winning neuron according to the weight update rule.

3. Defining the Main Function

In this section, we define the training data and initialize the weights. We also specify the number of epochs and the learning rate.

• T: This is the training data with four examples, each having four features.
• weights: These are the initial weights for two clusters, each with four features.
• epochs: Specifies the number of iterations for training.
• alpha: The learning rate used for updating weights.

4. Training the SOM Network

Here, we loop through each training example for the specified number of epochs, compute the winning cluster and update the weights. For each epoch and training sample we:

• Compute the winning cluster using the winner() method.
• Update the winning cluster’s weights using the update() method.

5. Classifying Test Sample

After training the SOM network, we use a test sample s and classify it into one of the clusters by computing which cluster has the closest weight vector to the input sample. Finally, we print the cluster assignment and the trained weights for each cluster.

6. Running the Main Function

The following block runs the main() function when the script is executed.

Output:

The output will display which cluster the test sample belongs to and the final trained weights of the clusters.

Test Sample s belongs to Cluster: 0
Trained weights: [[0.6000000000000001, 0.8, 0.5, 0.9], [0.3333984375, 0.0666015625, 0.7, 0.3]]

• Cluster 0 is assigned to the test sample [0, 0, 0, 1].
• The trained weights for both clusters are displayed after 3 epochs.