Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
There has been many theoretical papers studying neural nets in the setting that they behave like kernels. This work shows a clear example of functions that 1) cannot be learned in kernel setting 2) a neural net can learn it efficiently. Even though limitations of kernel methods are known among practitioners, this result is significant as it characterizes these limitations in a provable way. Therefore, I recommend acceptance.