In recent years, deep learning has been growing. However, if it is hard to imagine that a downturn in the sector could occur, this may well be the case according to a group of researchers. It would become increasingly difficult to make progress on this kind of artificial intelligence.
A limit soon reached?
Applications, chatbots, voice and facial recognition, object detection and manipulation, there are many innovations that integrate deep learning. At the end of 2019, for example, a South Korean researcher developed a method for predicting the El Nino weather phenomenon. However, this method is based on algorithms using networks of artificial neurons using deep learning.
Remember that this is a type of artificial intelligence (AI) derived from machine learning where the machine is able to learn on its own. The fact is that the growth of the sector is such that imagining a slowdown seems quite difficult.
And yet, a publication on the arXiv platform of July 10, 2020 believes that this could soon be the case. Researchers at MIT, MIT-IBM Watson AI Lab and Underwood International College say they have audited more than 1,000 articles on deep learning. According to them, making progress on this type of artificial intelligence will soon become very difficult, both economically and environmentally. For the authors of the publication, the costs will become so great that it will no longer be cost-effective to conduct this research.
Limiting costs and carbon footprint
However, the researchers predict that the problem could be solved with better algorithms and innovative machine learning methods. This could be particularly the case thanks to advances in quantum computing..
In addition, researchers from Intel and the University of Southern California (USA) may have already found a solution in 2016. They had used simple university laboratory equipment to train deep-reinforcing learning algorithms (DLR). For example, a single workstation allowed them to train a deepMind (Google) 3D AI to practice video game Doom.
An article published by IEEE Spectrum on July 17, 2020 gave the floor to Peter Stone, an AI expert at the University of Texas at Austin. He believes that making algorithms on basic material is a great research objective. According to him, this kind of project is beneficial in a complex energy and ecological context. Indeed, the resources usually required for this type of research are such that they have a very large carbon footprint.