A Lucky Catch: How a Coding Bot Built with Neural Nets Catch Programmers’ Thoughts


In this lightning talk, I will introduce a coding bot that I built together with Prof. Sabine Brunswicker, my mentor in Purdue University and PhD advisor to-be, where functions predictions in our recommender system is achieved by using neural nets on sentiment classifier to classify data. Inspired by the eminent neural conversational model built by Oriol Vinyals and Quoc Le in Cornell University, we incorporated the ‘sequence to sequence’ model when training our coding bot. Exploiting properties of neural nets, our new approach provides guarantees on the accuracy of the results as neural nets don’t analyze text, which is the functions wrote by programmers, at face value unlike the traditional approach that dissects the texts into smaller tokens, counts the number of times each token shows up and utilizes the ‘bag of words’ model, which is commonly called the ‘dictionary’ to compute the overall subjectivity of the text with the subjectivity of each word pre-recorded. Our new approach, however, will train the classifier and the given new function will be classified in a different category. Based on the theory of ‘sequence to sequence’ model, I will introduce the concept of representing different recurrent neural network in function predictions with different sets of parameters each, where the function is given few data points to learn from the hidden previous function and also the current function. Our model not only learns from the data injected by programmers but how it learns before. When predicting sequences of functions by converting text to vector or numerical representations, we not only predict based on the previous functions but all the functions that came before it by estimating the probability of certain contexts and learning a similarity function where the reply to the programmer is one of the elements in a predefined pool of possible answers. This ability to mimic programmers’ thoughts lead to a more efficient coding process where programmers just have to describe what they want to do to the coding bot to have the relevant functions inserted for them into their software applications.