Visual Editors for Machine Learning and Deep Learning are the hottest trend right now. Many people believe that it's going to be THE thing of 2020. But there is an issue with it.
I'm a big fan of democratising access to "gated" tech, skills and resources. The no-code movement, even if I personally don't prefer it over code-built, is a great thing that gave many creators the tools to do awesome things.
This trend extends to Machine Learning too. Recently, tools like Lobe.ai popped up (and quickly got acquired - damn!) and Apple also created their own system called CreateML with the goal to replace (or at least democratise) the building of Neural Networks with current frameworks, such as Keras or TensorFlow.
There is just one problem with these approaches: The difficulty of building Machine Learning models is not the coding, but the feature engineering, architecture and testing. While democratising access to the building itself is great, it almost makes it too easy to make a half-baked model public.
If you've heard about current algorithm bias and ethics problems, you know why this is so dangerous. Because models can be built and exported by everyone, it means that there is an increased danger of generating biased algorithms: Predictors that are pointing out certain ethnicities, classifiers that discriminate between genders, and possibly worse.
This is something that can be counteracted by creating a more assisted no-code tool, but the leading ones at this point only help with the actual implementation and training of a network, not with the feature engineering, data analysis or testing. The result will be biased and potentially dangerous network.
Also a big disclaimer to my no-code machine learning friends: This is in no way something that's unique to no-code tools. Code-based Machine Learning engineers create biased networks too. The issue is that these no-code tools are being sold to engineers who don't really have a basic foundation about Machine Learning. I've heard from plenty of iOS engineers now who pulled in CreateML into their projects, filled it up with 50-100 images that they made themselves, and got a trained model that they deployed in their app. That's now how it (usually) works and how you get something that's biased and mostly inaccurate.
I'm excited about a time where I don't have to touch TensorFlow anymore to be honest. The idea of clicking together my data, models and evaluation at the same quality than my code-based trainings achieve has a certain soothing feeling to it. Until then, use them with a grain of salt.