Supervised learning is the basis of deep learning. But, human and animal learning are unsupervised. In order to make deep learning more effective in human life we need to discover approaches using Deep learning to handle unsupervised learning. How much of progress is made in this direction so far?
With Deep Learning one can leverage unlabeled documents (unsupervised data) to have access to a much larger amount of input data, using a smaller amount of supervised data to improve the data representations and make them more related to the specific learning and inference tasks
Deep learning can be either supervised or unsupervised. It all depends on how we want to use it. Deep learning is used for highly scalable datasets and algorithms which are complex for machine learning algorithms like SVM, Logistic Regression e.t.c, to model. For example, Convolutional Neural Network (CNN) can be both supervised or unsupervised. If we want to classify images, we need to add dense layers and for classification, the training will be supervised. But, if we want to cluster images based on similarities of a group of images, we will extract the features using CNN and then use an unsupervised method like k-means or DBSCAN clustering.
In fact iterative techniques with convolutions are fundamentals for training data is an application of machine learning which is now a days called as deep learning. It may be parametric or non-parametric. Also it may be valid to supervised or non supervised data, however., it needs large sample data or big data. As for as humanise learning is concerned it is almost supervised learning with thinking.
Generative Adversarial Networks (GAN) is one of the examples of unsupervised learning techniques. It consists of a Generator (that will generate data) and discrimination (to identify between false and true data). During training these two models will work against reach the other two trains themselves and reach a point of equilibrium. Given a Gaussian noise as input to the Generator, it will produce some useful information. There are many variants of GAN available and they have applications in the image to image translation, generation of a fake image, etc.
Since to train a deep learning model usually a large amount of label data is required and that is usually hard to obtain. Therefore one can go for an unsupervised form of training.
Unsupervised learning is a machine learning technique, where you do not need to supervise the model. Instead, you need to allow the model to work on its own to discover information. Its important to gain insight into the way the system behaves without supervision. For example, GAN, CNN, K-NN, etc. Unsupervised learning algorithms allows you to perform more complex processing tasks compared to supervised learning.
Labelling data is a continuing hurdle in the application of supervised learning vis-a-vis cost, time and accuracy. If deep learning is to become widely applicable (just like software development is), it needs to get past this hurdle. Self-supervised / unsupervised learning (along with attention) are potential solutions to this problem. Ultimately, time will tell ...
I think the statement "human and animal learning are unsupervised" is not exactly correct, we do need to labeled data during our learning. Like you constantly correct your children when they learn at a young age. The limitation of unsupervised learning is not just labeled data, also the difficulty to convert abstract tasks in a machine-readable format.
Currently, there are some progressed, such self-supervised, using the property of data as pseudo-labels, Semi-supervised learning, etc.
Deep learning can be referred to as researchers, scientists, and/or the writer, who are devoted to knowing something beyond existence. With that kind of learning activity, you don't need a supervisor to tell you what needs to be done, but you need an editor to make your write-up okay.
- Badges
- Science topic