What is the current status of Deep Learning in economics research?

Question

Is Deep Learning currently being utilized in economics research, and if so, to what extent?

What is the future of Deep Learning in the economics profession?

Answer 1

I can't think of any papers using deep learning. That doesn't mean they don't exist but the large sample sizes needed for training are hard to come by. For applications, there are recent general machine learning papers by Susan Athey and by Sendhil Mullainathan that you should look at. Also look at Victor Chernozhukov's recent research.

Answer 2

This paper uses neural networks as a functional approximater.  He did decently on the job market with it:

http://victorduarte.mit.edu/about

IMO the use of neural networks as functional approximates when you have no interest (or there is no competing way) to understand why x causes y by looking at the parameters, is probably the best way to use ANN's in econ.

Comments

I should add this paper by Blei and Athey: https://www.youtube.com/watch?v=zwcjJQoK8_Q uses a Bayesian model very much based on word2vec which is a shallow Neural Network model.  Other mentions of word2vec include survey papers by Genzkow, Taddy and Kelly who obviously use this stuff for language modeling.  See here: https://web.stanford.edu/~gentzkow/research/text-as-data.pdf

Answer 3

My forthcoming NBER book chapter http://www.nber.org/chapters/c14009.pdf  has some overview of how machine learning is being used in economics broadly as well as predictions for the future (I'm predicting much greater adoption), with a particular focus on causal inference (and I also refer the reader to the JEP articles by Varian; Mullainathan and Speiss; and Belloni, Chernozhukov and Hansen).   

For some specific examples using neural nets, probably my favorite in economics that takes the idea beyond off-the-shelf prediction is this one by Greg Lewis, Matt Taddy and coauthors on "Deep IV": http://proceedings.mlr.press/v70/hartford17a.html

I also have several recent papers using neural nets embedded in more complex algorithms.
I have one on stable predictions (in KDD 2018, a top machine learning outlet) where we build on causal inference methods to find methods for prediction that are stable across environments.  Neural nets are incorporated for dimension reduction: https://arxiv.org/abs/1806.06270

I have two other papers in NIPS 2017 (another top machine learning outlet) that are using neural nets in combination with word embedding techniques, with applications to shopping data with many products.   http://papers.nips.cc/paper/7067-context-selection-for-embedding-models.pdf ; https://papers.nips.cc/paper/6629-structured-embedding-models-for-grouped-data.pdf

Comments

Quote:

I also have several recent papers using neural nets embedded in more complex algorithms.
I have one on stable predictions (in KDD 2018, a top machine learning outlet) where we build on causal inference methods to find methods for prediction that are stable across environments.  Neural nets are incorporated for dimension reduction: https://arxiv.org/abs/1806.06270

End Quote

I wasn't aware of this paper.  It is quite interesting (having only read the abstract).  Has anyone told you it is reminiscent of  Domain Adversarial Neural Networks (DANNs)?