What is Deep Learning?

Deep Learning was introduced by Dr. Rina Dechter to the machine learning community. She is professor of Computer Science at University of California, Irvine. Deep Learning is in fact a machine learning technique where computer architecture and algorithm draw inspiration from the structure and functions of human brain. Using Deep Learning computer model performs different types of classification tasks directly from images, text or sound and that too with accuracy same as that of human minds. Deep learning models are based on communication process and data processing biological nervous system. In this the basic parameters are set up about the data and teach the computer to recognize patterns by using multiple layers.

(Source Ref. 2)

Deep Learning as we now know is sub-field of machine learning. Let us understand how and where exactly it discretely fits into the machine learning field. In machine learning, the learning process is supervised and the programmer needs to be accurate when telling the system the type of thing it should be looking for when deciding that an image contains for example a bird or not. This hectic task is called feature extraction and the accuracy depends upon programmer's ability to accurately define a feature set of bird. The benefit of deep learning is that the program builds the feature set by itself without supervision. Unsupervised learning is more accurate and faster. Figure below shows a simple example of machine learning and Deep Learning model.

Figure 1: Machine Learning and Deep Learning Model (Source: Ref. 3)

Machine learning works with manually extracted relevant feature from the images and categorization model is made using the feature extracted. With Deep Learning relevant feature is automatically extracted from the images and end to end learning process takes place. Conventional machine learning programs analyze data in linear form while Deep Learning uses nonlinear way to process the data - it uses neural network architecture and therefore sometimes referred to as deep neural networks model. Deep Learning is far more capable when compared to machine learning as each algorithm used in Deep Learning applies non linear transformation on its input and creates a statistical structure as output.

The main advantage of Deep Learning is that it continues to improve with increase in size of data. Deep Learning algorithm can take non-categorized and messed up data like images, audio recordings, videos etc., it can test and implement enough order on the data to make it useful for prediction, and developing a structure of feature like for example rat or bird in an image, sound that form a word in speech etc. Deep Learning is used in NLP processing, image recognition tools, bioinformatics and speech recognition software. It can also be used to restore the color of black and white picture using a system known as let there be a color.

In another decade or so Deep Learning systems, languages and libraries will become the base of every software development toolkit. In near future we will have Deep Learning applications in areas like automotive (self-driving car), health care (neural network for brain cancer detection etc.), automatic handwriting generation, voice control in consumer devices and many more.

By: Dr. Sheifali Gupta (Professor), Dr. Shalli Rani (Associate Professor) and
Ms. Urvashi Sharma (Research Scholar), CURIN, Chitkara University

References

1. https://in.mathworks.com/discovery/deep-learning.html
2. https://aitrends.com/deep-learning/here-are-22-selected-top-papers-on-deep-learning/
3. https://medium.com/swlh/ill-tell-you-why-deep-learning-is-so-popular-and-in-demand-5aca72628780