The Two Cultures

This started because I was listening to the ‘Subject to:’ podcast.

I was listening to Dimitris Bertsimas speak about the state of modelling and he suddenly brings up this article by Leo Breiman, called “The Two Cities”. Well, I later found out it was called ‘The Two Cultures’ and decided to give it a read since Dimitris recommended us young listeners to.

Going through it I was immediately struck with this chart:

Alright, so nature has a way of converting some inputs to some form of output. (Think, I eat food and tomorrow it is transformed into some other format.)

So how can we understand such processes?

Breiman states that there are these ‘Two Cultures’

Initially looking at these charts, I thought, aren’t they just the same thing?

But the key distinction is in the labelling of what’s in the box.

In the Data model, we attempt to populate it with our own models. Believing that we are able to create some sort of understanding of natures mechanisms.

Whereas in the Algorithmic model, we treat nature as something we don’t fully know and just bypass the modelling of it, aiming to go from x to y directly.

As we advocate for more parameters and deeper machine learning models, I have been thinking that it hasn’t been the best paradigm due to the lack of understanding of the models inner workings.

The other day I also saw this LinkedIn post with this meme:

It made me think, is it really that bad that we have adopted this paradigm of more data = better?

As long as we can get from x to y accurately enough, does it really matter?

I suppose eventually we will have to shift back to a more data modelling way. A means to interpret all this algorithms we have created.

But for now, I don’t know if interpretability matters, as long as we keep seeing the results.