DataScienceToday is a new online plateform and going source for data science related content for its influential audience in the world . You can reach us via email
contact@datasciencetoday.net
datasciencetoday.net@gmail.com
made by Bouchra ZEITANE , Zaineb MOUNIR
under the supervision of Pr. Habib BENLAHMAR
Unsupervised learning works by analyzing the data without its labels for the hidden structures within it, and through determining the correlations, and for features that actually correlate two data items. It is being used for clustering, dimensionality reduction, feature learning, density estimation, etc.
The hidden structure sometimes called feature vector, represents the input data such a way that if the same feature vector is being use to reconstruct the input, then one can do that with some acceptable loss. The variance in two feature vectors of two inputs directly proportional to the variance in the inputs itself. Thus this hidden structure or feature vector only represents features in the data that actually give distinction to it.
RBM, autoencoders are the two simple form of unsupervised neural networks. Moreover a CNN network without a FC network can be used as a encoder for the images. The training for such encoder networks are done by using a decoder network, and optimizing by reducing the reconstruction loss.Because there is no external teacher in unsupervised learning, it is crucial to increase the entropy which can be done by redundancies in the data. Redundancy provides knowledge for unsupervised learning. Unsupervised learning works with the mechanism that compare the coming data with the datas seen before.
What unsupervised learning models actually do is to measure the familiarity of coming datas with the past seen datas, and make inferences with that comparison like clustering.Unsupervised learning often tries to take advantage of statistical patterns that reoccur in data. Since unsupervised learning generally does not have labels to work with, the algorithms have to do the next best thing which is try to figure out what commonly (i.e. often repeatedly) happens in data and compare that against what uncommonly happens in data.