Hello to all
Artificial intelligence is often mentioned as an area where corporations make large investments. It seems that they use AI in autonomous vehicles, face recognition, medicine, banking, financing, weather forecast, etc.
The celebrated backpropagation algorithm has wide use within academic circles. Many papers exist that discuss, analyze and apply BP. But is it actually used in any of the practical developments undertaken by large companies? If so, in which specific ones. If not, which is the preferred NN training method for these commercially oriented applications?
With cordial regards,
Daniel Crespin
It seems backpropagation is used in SAS, a traditional commercial software.
https://communities.sas.com/t5/SAS-Communities-Library/How-to-build-a-deep-learning-model-in-SAS-Enterprise-Miner/ta-p/231190
Eduardo Corrêa Gonçalves
Thanks very much Eduardo for the link to the SAS backpropagation package. There is no question that the SAS software uses BP in the sense that it provides language and underlying machine code to train NNs.
But I want to know if there is some outstanding practical use of BP in applied artificial intelligence. In the particular case of the SAS BP, has it been employed to train a NN that is presently used in an important task?
For example, at academic and research levels there has been much activity on reliable computer systems that detect cancer in a mammography. My question would then be if some system that effectively performs this detection has been developed by means of BP and is today in actual clinical use.
There is the face recognition problem for which artificial intelligence seems to be presently used. Are their algorithms based on BP training of NNs? Perhaps they only require appropriate use of a data basis plus traditional statistical tools.
With cordial regards,
Daniel Crespin
Yes, there are a considerable number of practical applications. You may, for instance, examine the following article:
- Yann LeCun, Yoshua Bengio & Geoffrey Hinton . "Deep Learning". Nature, v. 521, 28 May 2015.
Back-propagation is a fancy name for error reduction techniques starting from circuit theory to control systems rooted in calculus . Now with stats estimation and convergence is extremely relevant and so is BP . Every domain where there is prediction, error control is needed ...
Eduardo Corrêa Gonçalves
Thanks very much
The paper by Yann LeCun pointed to by Eduardo sufficiently answers my question. It is now clear that BP has not been replaced and remains important down to practical real world applications. From LeCun paper and after some navigation I found the following additional references:
Ian Goodfellow, I. et al Deep Learning (book by MIT Press)
Jürgen Schmidhuber, Deep Learning in Neural Networks. An Overview
Cireçan, D. et al - Flexible, High Performance Convolutional Neural Networks for Image Classification
Time ago I published the paper Generalized Backpropagation containing a deduction of general BP formulas
https://www.researchgate.net/profile/Daniel_Crespin/publication/236214615
and with BP still useful may now write a further generalization.
Hearty thanks to all, especially to Eduardo.
Daniel Crespin
- Badges
- Science topic