DeepFix: Fixing Common C Language Errors by Deep Learning
It is essensically a multi-layered sequence-to-sequence NN.
couple of keywords
encoder-decoder framework & sequence-to-sequence problem: a basic sequence-to-sequence model has 2 rnn:
- encdoer: process input
- decoder: generate output sequence the encoder and decoder are trained together.
slacked gated recurrent unit energy value context vector soft alightment model
== define a fixed size pool of name construct encoding map mark the line number output: == a sequence of tokens == Sequence-to-sequence model encoder map token to a real vector(annotation) context vector: weighted sum of input sequenced annotations with normalized weight energy and weight determines importance of annotation decoder output result
Program predict iteratively to repair multiple errors
Random thoughts: different token, then after repair, it repair the same mistake over and over
mutate up to 5 state from each correct program to introduce errors
How does oracle decide
Paper to read