Human-Learn: Draw Machine Learning Models | Hacker News Reader

Human-Learn: Draw Machine Learning Models

koaning.github.io

53 pointspolm239 comments

This is a really interesting project. It's a collection of utilities for constructing rule-based models, as opposed to statistical models (i.e. ML).

I can see myself using it for exploratory analysis and constructing baseline models.

As for the "let's just draw the model" idea, it gets complicated as soon as you're dealing with more than one input parameter. The author has some workarounds, it looks like:

https://koaning.github.io/human-learn/guide/drawing-classifi...

Still not sure why this repo doesn't implement dimensionality reduction techiques. Seems like a huge oversight since that's what you need for doing this on high dimensional data.

"Let's draw the model" works in a PCA or UMAP space if your boundaries are already pretty good. I'm surprised that we haven't seen more machine teaching tools like this...

This is a neat demo, but at the same time, what any of the more sophisticated ML and deep learning techniques do is effectively create that picture for you. If you're at the stage of seeing the data like that, the hard work has been done. Being able to project the data into a latent space where classes can be separated is what makes modern ML so valuable.

PCA is so fast that it can run on 100 million data-points and give you results in minutes with the proper implementation of it.

If you can find a way to project that new data quickly, and if the decision boundries are instantly recognizable to us humans - one can draw the boundary faster than a model could generate it given the data. The human drawn boundary has bonus regularization built right in too (because "sub-optimal" boundaries on the training data may generalize better)

I compared human models to xgboost models. The latter were much more complex, completely incomprehensible to humans, but the former were simple and understandable. I think human generated models is a promotion sing and significantly under explored area.

If your boundaries are so easily visualized that you can draw a circle around them, then why not use a fully automatic(tm) decision tree, and let the computer do the hard work?

This is like the fits people drew on paper before computers were available.

humans provide accidental regularization - and a human made decision boundary can have further fine-tuning by an expert. This is important related to things like calibrating probabilities, which decision trees do very poorly on even if they get good scores on the metric they're trained on.

Even better if you have a generative model which can "hallucinate" parts of the decision boundary which don't have points. Now you can "probe" your model and have a human intervene if they think parts of the decision boundary are wrong

I suppose a human can modify a decision tree as well - but in practice human generated rules may be better for some domains.

Absolutely brilliant. Ever since taking ML classes in MSc and then again in PhD I have wondered why no machine learning approach ever applies such an obvious technique. And most real world datasets are probably adequately addressed with this approach, and everyone regardless of technical background can understand it. ML no longer has to be an inscrutable black box mystery! My hat is off to you sir!

This sounds neat as a type of low code way to make Machine Learning models but how much different is this than using pandas?

Reply on news.ycombinator.com