How and why did this end up on the homepage again?[a][b]
This is a silly claim, backed only by tiny toy experiments on tiny toy datasets with tiny toy neural networks. See links to the paper in [a] and [b] below.
Using the same logic as the authors of the paper, I could make the claim that neural nets are "essentially the same" as any universal approximator (GP's, SVM's, RF's, GLM's, etc.) and then run a few experiments only on tiny toy problems to back up my claim.
The authors should try to fit a polynomial regression model to cognitive tasks such as language translation, image generation, or modeling the game of Go... they won't get very far without running into the curse of dimensionality.
Please stop posting this on HN.