"Deep learning is dead, differential programming is long live" LeCun teacher responds

Deep learning est mort. Vive differentiable programming!

This English-French mixed words, translated into Chinese, is "deep learning is dead, can be differential programming long live." It is one of the big three in deep learning: Yann LeCun.

Nani.

Whether it was Ali Rahimi's "Alchemy said" last month, or the previous days Gary Marcus questioned the depth of learning, LeCun teachers have come forward to tough back to godless, today.

Well...... In an article on Facebook, LeCun said that deep learning had died before it was dead, saying "deep learning has been too long to be used as a buzzword." ”

Just another way of saying it. What is "differential programming".

LeCun admits that "differential programming" is simply a renaming of the modern deep learning technology system, similar to the "deep learning" variant of the neural network that had more than two hidden layers in that year.

He also said:

But the point is that people are now building a new class of software by assembling parametric functional module networks and training them with some sort of gradient-based optimization.

A growing number of people are programmatically defining networks in a data-dependent way (using loops and conditions) to change as the input data changes dynamically. In addition to parameterization, automatic differentiation, and the training/optimization features, this is much like a normal program.

Dynamic networks have become increasingly popular (especially for NLP) thanks to deep learning frameworks such as Pytorch and Chainer (note: The previous deep learning framework lush can handle a special dynamic network called graph Transformer Networks , for text recognition).

Now, people are actively studying the compiler of the command-type differential programming language, which is a very exciting way to develop learning-based AI.

In addition to the thought that deep learning should be called differential programming, LeCun also said he decided to his recently particularly respected "predictive learning", renamed "Imputative Learning", we temporarily called "Attribution learning" good.

He finally made a small announcement that he would talk more later, but the quantum bit didn't wait until noon in the morning.

Had to keep watching his Facebook together.

Https://www.facebook.com/yann.lecun

For the new idea of LeCun, many people on the foreign forums are too long and too difficult to call the new name, but some people say that "programming" is indeed more appropriate than "learning".

--------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------