This paper presents analysis of what 3-layer neural nets with non-smooth activation functions can learn. Extending such analysis to three layers is very challenging and all reviewers find this result interesting. Therefore, I recommend acceptance.