Unsupervised Learning – A New Era of Machine Learning
One of the most remarkable achievements of the 21st century has undoubtedly been the advancement of neural networks. The development of neural networks has been so rapid that they are being used for everything from detecting small cells to treating cancers to driving cars all without any human intervention. This has become possible because of rapid advancements in computational power and the development of GPUs.
Expectedly, there are various problems which hinder further development of neural networks. Most of the machine learning and deep learning approaches currently revolve around supervised learning with differentiable layers. Though the success of supervised learning models has been great, they fail in cross-domain performance. Suppose, for example, that a particular model has been trained for an image classification task, but we want the same model to count the objects in the image which are useful in an image captioning task. Current models will fail miserably. This is a major problem: supervised learning lacks generalizability and robustness.
The current necessity is to make a model trained for a particular task work well for different tasks. This is where unsupervised learning comes to the rescue. The biggest advantage with unsupervised learning is that it is successful in extracting useful information from high dimensional data. This extracted information can further be used for doing various tasks. So before going into the prediction part, the first task is to know whether there are any patterns among the classes to differentiate within. Here come two algorithms to the rescue: t-SNE (T Distributed Stochastic Neighbour Embedding) and MDS (multidimensional scaling), which help visualize the high dimensional data.
The approach of t-SNE is very simple. It first converts similarities between high dimensional data points to joint probabilities, and then it further tries to minimize the Kullback-Leibler divergence between the joint probabilities of the low-dimensional and the high-dimensional data. MDS also does the same thing. The only difference is that MDS successfully preserves distances and global structures but t-SNE cannot.
The above figure shows the t-SNE visualization of the MNIST dataset. Thus, each class can be segregated based on the visualization. The t-SNE helps conclude further classifications of tasks on the dataset
In the next blog, further predictions will be discussed.
Written by Pranav Agarwal, Edited by Jack Vasquez & Alexander Fleiss