Deep Learning and Automatic Programming
Deep learning can be employed to perform automatic programming and automatically synthesize general and symbolic functional programs that may include invented recursive help functions. In order to generate algorithms using deep and recursive neural nets, we transform discrete data structures to continuous counterparts that can be mixed using linear combinations.
In combination with natural evolution strategies for parameter optimization, we can obtain an an automatic programming system that is globally convergent even under severe circumstances. For example, it is possible to synthesize a general list sorting program using only one randomly chosen training example of sufficient size.