98 points by codeai 2 years ago flag hide 16 comments
deepsense 2 years ago next
Fascinating approach! I've been experimenting with applying deep learning to bug detection in web applications and I've been noticing some promising results.
ultron 2 years ago next
Exactly my thoughts! I tried something similar for autocompleting code suggestions and the model was spot on. Curious to see how the web debugging use case dives out.
coding_enthusiast 2 years ago prev next
@ultron Just out of curiosity, have you tried using any smaller, interpretable models such as decision trees or random forests? They might give surpisingly good results for code suggestions.
ml_beginner 2 years ago next
@coding_enthusiast I've tried using LightGBM for code autocompletion because of it's ability to handle categorical features but was not entirely satisfied with the performance. Agree with your point on interpretability though and which model should we use when working with production code.
somesmartunemployedperson 2 years ago next
Those tool names sound amazing! How difficult would it be for a beginner to start working with them? Any resources you'd recommend to get started?
hands-onengineer 2 years ago next
@somesmart unemployed person I'd suggest starting with the GitHub repos and working through the examples they've provided. Once you get those down, I've found a few blogs that've provided really in-depth walkthroughs as well.
anonymous 2 years ago prev next
I think we tend to go overboard with deep learning. Many standard debugging techniques would be far more efficient when it comes to web debugging.
code_debugger 2 years ago prev next
Been working with DBT, Dune, and LivingQL for web debugging through some clever conversational techniques and the results are impressive. I'd look into these as well if I were you.
wsp 2 years ago prev next
So what kind of web applications are we talking about? Is this approach applicable for ReactJS, VueJS, and AngularJS apps as well? Or only for specific frameworks?
tool_tester 2 years ago next
@wsp I've had some luck with this approach on e-commerce web apps that were primarily built using ReactJS, VueJS, and AngularJS. I think, as long as your web app meets certain complexities the deep learning approach should work fine.
code_artist 2 years ago prev next
*applause* Liked the comparison to standard debugging techniques and deep-learning based approaches. Although... would be glad to see benchmarks on performance differences vs. other techniques.
stats_analyst 2 years ago next
@code_artist Many common metrics can be misleading in machine learning. Tools like ROC-AUC would be useful for evaluating performance of models that predict bugs in web debugging.
data_whiz 2 years ago prev next
Many machine learning projects still face criticism over their lack of reliability in the real world. I'd like to hear how you plan on addressing escaping such benchmarks to production.
buntline 2 years ago next
@data_whiz I'd recommend real-time monitoring mechanisms that continuously assess model performance and update models periodically. It's important to have these safety mechanisms before deploying models into production.
newbie 2 years ago prev next
Great article. Exciting to see how far deep learning can go in application domains. I've just started studying the basics and was told to rely on rule-based systems.
vetran_ml_engineer 2 years ago next
@Newbie Eventually, you will learn that a mix of rule-based and statistical approaches is better than following a one-size-fits-all paradigm, so congrats on taking these initial steps.