Did we overfit taste?

[August 17, 2024]

Well, yeah, I think so. Here, Let me explain.

Did you ever think about the irony of our palates? How can the food that tastes really good be almost certainly considered bad for our health? Isn't the core function of our taste buds to prevent us from consuming things that are "not good for our health"

As far as I have read/understood, the answer lies in using a bad proxy metric!

Yeah, we use taste as a proxy metric for health. Sure, it made sense some couple of 100 years ago, when we were natively attracted to sugars, fats, and salts, given these are really important nutrients for our body. Our body was tuned to crave it and was drawn to it for sustainability (and a sensible diet).

This harmony was broken when we gained the ability to change/modify the foods that are available to us. It broke the fundamental relationship. We can now add fats, and sugars in amounts to food way beyond what is good for our health, and eat those exclusively, rather than eat with a balanced diet of fibers, grains, and meat that made a sensible diet historically.

And the more skilfully we can manipulate food (and the more our lifestyle diverges from our ancestors), the more imperfect a metric "taste" becomes.

So yeah. we did overfit "taste", and hence, it is no longer a reliable metric to predict what is healthy for us.

Thus,

Overfitting, fundamentally, is a consequence of focussing on what we've been able to measure rather than what matters, like idolatry of data."

-Brian Christian

What is the point of this post?

Well, concepts in machine learning extend to life, or rather life concepts have inspired machine learning as a science. This was an example where a proxy metric has gone awry leading to a hell lot of problems.

A piece of advice after having over a decade of experience in building intelligent products, is don't lose focus on what really matters.

The model is just a means to an end. Sure, objective metrics have their place in the development lifecycle but to rely solely on them and discard everything else as someone else's problem does not create real value.

The model metrics are proxy metrics (like *taste*) that capture the notion of good under certain assumptions and go awry when the relationship changes.

So, be wary of what you can measure and what really matters.

Be okay(in fact be curious and embrace) stepping out of the "modeling" stage for that is where real learning lies.