Deep-learning artificial neural networks

Developers

J. Holfold, D. E. Rumelhart, S. I. Bartsev, v. A. Okhonin, Y.LeCun etc.

Description of the technology

Artificial neural network is a mathematical model and its program or hardware realization, created on the principles of organization and functioning of biological neural networks, i.e. networks of nerve cells of the living organism.

Neural networks are not being programmed in the common sense of that word, they are being learnt. The possibility of learning is one of the major advantages of neural networks against traditional algorithms. There are various methods of learning. While solving thousands and thousands of training tasks, a neural network learns to find complex interconnections between input and output data and make general conclusions. Interconnections found by the neural network are fixed by the intensity of the connections (weighting factor) between the neurons. Human brain functions in a similar way (here, agitation strength of the living neurons plays the role of the weighting factor).

Deep neural network (DNN) is an artificial neural network that is more than two-layer deep. It is believed that every next layer learns a new level of data abstraction.

Nowadays, deep neural networks, or deep-learning artificial neural networks is a rapidly growing branch of knowledge which is connected to an artificial intellect and which is a catalyst for a wide variety of industries, from medicine and pharmacology to autopilot cars.

Practical application

This technology will have a great future in various areas, including medicine, due to the possibility to teach neural networks to search for the optimal therapy. When sufficient amount of data (including data about cancer and aging) is accumulated and those data are treated in a specific manner, it will be able to «train» deep neural networks to search for the optimal therapy and for the new drugs. The method of deep learning will obviously be used in the development of effective drug combinations, which can delay age-related pathologies or even repair damages. The long-run objective is a victory over aging.

In medicine, this technology is at the stage of preclinical and clinical trials.

Laboratories

• Deep Knowledge Ventures (Hong Kong)
• InSilicoMedicine (USA)
• ПОНКЦ (Russia)
• NVIDIA (USA)
• IDSIA (Dalle Molle Insliiule for Artificial Intelligence) (Switzerland)
• Culture Lab, Newcastle University (UK)
• IDEA Lab, Biomedical Research Imaging Center, UNC School of Medicine (USA)

Links

Publications

• Hinton, Geoffrey E., Simon Osindero, and Yee-Whye Teh. «A fast learning algorithm for deep belief nets." Neural computation 18.7 (2006): 1527–1554.
• Hinton, Geoffrey E. «Learning multiple layers of representation." Trends in cognitive sciences 11.10 (2007): 428–434.
• Cireşan, Dan C., et al. «Mitosis detection in breast cancer histology images with deep neural networks." Medical Image Computing and Computer-Assisted Intervention-MICCAI 2013. Springer Berlin Heidelberg, 2013. 411–418.
• Hammerla, Nils Yannick, et al. «PD Disease State Assessment in Naturalistic Environments Using Deep Learning." AAAI. 2015.
• Suk, Heung-Il, Dinggang Shen, and Alzheimer’s Disease Neuroimaging Initiative. «Deep Learning in Diagnosis of Brain Disorders." Recent Progress in Brain and Cognitive Engineering. Springer Netherlands, 2015. 203–213.
• Hahm, Seongjun, and Jun Wang. «Silent speech recognition from articulatory movements using deep neural network." Proc. of the International Congress of Phonetic Sciences. 2015.
• Барцев, С. И., В. А. Охонин. «Адаптивные сети обработки информации." Красноярск: Ин-т физики СО АН СССР (1986).
• LeCun, Yann, et al. «Backpropagation applied to handwritten zip code recognition." Neural computation 1.4 (1989): 541–551.