A brief introduction to Unsupervised Learning

A space in machine learning which is evolving as time passes from east to west

To begin with, we should know that machine primarily consists of four major domain.

1. Supervised learning: An agent or algorithm learns from the labeled data.
2. Unsupervised learning: An agent or algorithm learns from the unlabeled data i.e. it finds similar patterns in the dataset and groups them accordingly.
3. Semi-supervised learning: A combination of both Supervised and Unsupervised learning.
4. Reinforcement learning: An agent or algorithm learns patterns or behaviors by correcting itself over and over again until evolves into a better agent.

Now let us see the methods which come under the unsupervised learning domain.

Clustering

Photo by Nareeta Martin on Unsplash

The goal of clustering is to create groups of data points such that points in different clusters are dissimilar while points within a cluster are similar.

Clustering also has its own subcategories.

1. K-means clustering

With k-means clustering, we want to cluster our data points into k groups. A larger k creates smaller groups with more granularity, a lower k means larger groups and less granularity. It can be compared to the separate crowds of people surrounding different famous people at a party. The density of the crowd depends on the fame of that person.

2. Hierarchical clustering

Hierarchical clustering is similar to regular clustering, but it focuses on building a hierarchy of clusters. This type of clustering is used in the online shopping websites, where there are broad categories for simple navigation displayed on the homepage and as you click on it, further specific categories related to that would be displayed. This explains the more distinct cluster of items.

Dimensionality-reduction

1. Principal Component Analysis:

PCA is a dimensionality-reduction method in unsupervised learning which is used to reduce the dimensionality of large data sets into smaller ones by choosing the basis vectors on our own which are known as principal components. PCA remaps the space in which our data exists to make it more compressible. The transformed dimension is smaller than the original dimension.

2. K-nearest neighbor

How do you determine the housing price of a house in a particular locality? We would take the average of the price of the houses in the nearby locality and determine the approximate price of the house we are about to buy. We label the test data point based on the average of the sample data in its neighborhood. We take the mean of the values if the variables are continuous and mode if they are categorical.

Applications of k-NN:

• Helps in the update of new methods of fraud detection.
• Determining the housing price and detection of the temperature in the locality.
• Imputing missing training data.

3. T-distributed Stochastic Neighbor Embedding

t-SNE Embedding is an algorithm used to reduce a high dimensional dataset into a low dimensional graph that retains most of the original information. It is based on the principle of determining the similarity of all points in the scatter plot.

The process done here is measuring the distance from the point we are interested in all the other points and plotting that distance on a normal distribution curve, which is centered on the point that we are interested in.

Note: We use a normal distribution curve because distant points have low similarity values and close points have high similarity values.

Now it puts the data points on a number line in a random order, and t-SNE moves these points little by little based on their similarity values, until it has clustered them properly on a lower dimension.

Generative modeling

1. Generative adversarial network

A generative adversarial network is deep learning-based generative model. Generative models are models that use unsupervised learning. GAN is a system where two neural networks compete to create or generate variations within a dataset.

It has a generator model and a discriminator model. The generator network takes a sample and generates a sample of data by learning the distribution of classes. The discriminator network learns the boundaries between those classes by estimating the probability of whether the sample is taken from the real sample.

Applications of GAN :

• They are used for image manipulation and generation.
• They can be deployed for tasks in understanding risk and recovery in healthcare.
• Used in drug research to produce new chemical structures from the existing ones.
• Google brain project is an interesting application of GAN.

The main advantage of GAN is to generate data when there is not much data available, without any human supervision.

2. Deep Convolutional Generative adversarial Network

DCGAN has convolutional layers between the input and the output image in the generator. And in the discriminator, it uses regular convolutional networks to classify the generated and the real images. The architecture of the DCGAN is:

• The pooling layers are replaced with generators and discriminators.
• Batch normalization is used in both generators and discriminators.
• The fully connected layers are removed.
• ReLU is used as the activation function in the generator for all layers except the output layer.
• Leaky ReLU activation function is used in the discriminator for all layers.

3. Style Transfer

Style transfer is the method used to generate a new image by combining the content image with a style image. By using this we can make the environment image that we have looked a lot greater because it is being combined with the style of iconic paintings.

The activations in the neural network of the content and the style image should match the activations in the generated image. So style transfer can make any image that you took on your trek look modified like the famous Hokusai Japanese painting.