Machine Learning technology empowers various aspects of modern society: right from web searches to content refining on social networks to recommendation systems on e-commerce websites, it is increasingly present in consumer regular products such as smartphones and cameras.
The systems built on Machine learning are widely used in identifying objects in images, transcribing speech into text, matching news items, postings or products with users’ interests, and selecting relevant results of search. These applications increasingly make use of a class of techniques called Deep learning.
Deep-learning methods are a kind of representation learning methods that allows machines to feed with raw data to discover the representations needed for classification and detection automatically.
These learning methods are built with multiple levels of representation obtained by composing simple, non-linear modules that transform the representation at each single level, firstly starting with the raw input to a representation at a higher abstract level.
With this composition of such transformations, very complex functions are processed to learn. Consequently, in case of classification tasks, higher layers of representation turn up aspects of the input that are important for discrimination and put down irrelevant variations.
The key aspect of deep learning is that these layers of features are not designed by the intervision of human engineers: they are meant to learn from data provided using a general-purpose learning and training procedure.
Deep learning is making significant advances in solving problems that have countered the best attempts of the artificial intelligence community for many years. It has turned out to be very good at discovering convoluted structures in high-dimensional data and is therefore made applicable to many domains such as science, business and government.
The new learning algorithms and its architectures which are being developed upon the deep neural network, tend to accelerate the progress in the near future. But, many of us think that deep learning will have success further as it requires very little engineering by hand, so that it can easily have the advantage of increasing computational data.
Human vision is an operational process that sequentially samples the optic array in an intelligent, problem-specific way using a tiny, high-resolution fovea surrounded by a low, large-resolution.
The future progress is expected in vision to come from systems that combine ConvNets with RNNs and train end-to-end that use reinforcement learning to decide to look after. Systems combining reinforcement learning and deep learning are in their origin, but they already outperformed passive vision systems at classification problems and produced impressive results in learning to play many different video games.
Natural language understanding(NLU) is another part in which deep learning is poised to make a greater impact over the next few years. Systems that use RNNs to understand whole documents or sentences would become a much better learning strategy to selectively attend in one part at a time.
Ultimately, vital progress in artificial intelligence will come up through systems that combine complex learning with representation learning. Although simple reasoning and deep learning have been used for handwriting recognition and speech recognition strategies for a long time, new paradigms are required to replace rule-based manipulation with symbolic expressions by operations on large vectors.