Find out how deep learning has become so important in recent years and the prospects for this learning technique.
About ten years ago, deep learning was still an obscure discipline that belonged to science fiction. Imagining that machines could learn on their own was a bit of a headache. Today, many artificial intelligence applications are based on deep learning. The boundary between human and machine continues to blur even if limitations persist.
What is deep learning?
Deep learning is a learning method for computer systems. The machine learns by itself by analyzing unstructured data through several layers of algorithmic models. Deep learning therefore consists in the creation of artificial neural networks and is similar in this sense to human learning.
Deep learning is part of machine learning techniques. It is commonly used in the context of image recognition or speech recognition. The market is currently experiencing a rapid rise. A report published by Research & Markets in April 2020 estimates its annual growth rate at 30% in the next five years.
History of deep learning
Talking about the history of deep learning may seem too froward to some. We are indeed only at the beginning. However, while many people have only recently become familiar with this technology, the theory itself goes back several decades.
As early as 1943, Warren McCulloch and Walter Pitts were talking about the concept of an artificial neural network. Thereafter, this concept was the subject of university research in the 1950s and 1960s, notably at Stanford, MIT and the University of Toronto.
However, it was not until the late 1990s that the technology received a real boost. In 1998, Frenchman Yann LeCun published the results of his ground-breaking research on the use of artificial neural networks for image recognition. Winner of the Turing Prize in 2019 with Geoffrey Hinton and Yoshua Bengio, he is considered one of the founding fathers of deep learning. Yann LeCun is now in charge of artificial intelligence for the Facebook company.
Of course, others have contributed in parallel to the emergence of the concept of deep learning. Names such as Frank Rosenblatt, John Hopfield, Andrew Ng or Ian Goodfellow have marked the evolution of this discipline and history continues to be written. Many researchers are actively working to make machine learning methods even more effective.
Why has deep learning revolutionized the field of artificial intelligence?
Previously, the machine learning process was done by training the machine and giving it a series of practical examples. The recurring example is that of the cat. The model trained by ingesting thousands of pictures of cats. It was then able to recognize a cat's photo. The data had to be structured and labeled.
From now on, learning goes further. The machine becomes able to analyze astronomical volumes of unstructured data and to identify itself the links or interactions between certain concepts thanks to the many layers of analysis it has at its disposal. The machine will therefore understand that a cat exists by locating a common core within the images or other data presented.
The exponential growth of the volume of data available now allows these concepts to become a reality. In order to guarantee the effectiveness of deep learning, the machine must have enough data to detect correct patterns (i.e., some kind of structure). This explains why this technique is currently experiencing unprecedented momentum.
The potential for deep learning is therefore enormous. On the one hand, deep learning allows human capacities to be exceeded. In image recognition, the best systems have been outperforming human performance since 2014 already. On the other hand, the predictive potential of deep learning is the object of all covetousness. In the financial sector in particular, ever more sophisticated models and intelligence seek to use data to predict course changes. The sums invested are obviously commensurate with the stakes involved. More simply, there are many applications in various fields such as weather or predictive maintenance.
The market potential is enormous. In 2018, McKinsey estimated the annual value creation between 3,500 and 5,800 billion dollars. The O'Reilly company indicates that unsupervised learning would have increased by 172% in 2019 and represented 22% of artificial intelligence applications. Beyond the economic outlook, potential opportunities in the medical field in particular could become essential. One thinks for example of diagnostics.
Are expectations too high in the face of deep learning?
While enthusiasm seems to be shared by many, some voices denounce excessive expectations in the face of deep learning. Others are concerned about the possible biases of an artificial intelligence that is out of control. Finally, some also believe that many deep learning techniques will be limited in the future by various regulations.
It is true that the technology is the subject of much ethical and regulatory debate. Respect for the principle of privacy also raises controversy.
Moreover, the concept of deep learning remains poorly understood. Definitions are multiplying and vary from one company to another. For many, deep learning is simply opposed to supervised learning modes, which are the default learning techniques. Some also distinguish between deep learning and artificial neural networks.
The absence of clear definitions shared by all is also cruelly lacking here, as is the case with many recent technologies. If you ask two people to define the scope of deep learning, the chance of getting an identical explanation is similar to winning the Euromillions.
Ryax and deep learning
As a data processing platform, Ryax allows companies to use their machine learning and deep learning algorithms through a unified framework. The use of software such as Ryax to facilitate the implementation in production has many advantages to optimize the exploitation of data. Using our drag & drop interface, the deployment of advanced data-science workflows becomes much simpler.
La Ryax Team.