Work Log Day 274

Since the last post, a new model has been trained that improves upon the previous. Achieving mid 70s for the MSE. However I decided recently to start pulling shots on the Niche Zero again, as the model has been trained largely on the MC3. Thus the model is back up in the mid 80s. The fact that both the Niche and the Flat caused the model to get worse, I worry that the model can only be trained effectively to a specific grinder. And so I end up in the position where I need to generate more data to train the model against. I have actually begun to wonder if ML is going to have a significant impact on generating better data, as data with a lot of variance is almost useless to train (unless you have massive quantities of data). So while I chug along collecting data my mind wanders to generic coffee subjects.

General Coffee Questions

Having been collecting TDS data on my espresso shots for awhile now I have increasingly questioned the notion that “higher TDS correlates with taste”. It might stem from this paper, which really claims that sensory properties change as a function of TDS. And even then, that was an evaluation of brewed coffee where very small changes in TDS might have a bigger impact on the overall beverage? Now I have also heard the caveat that higher TDS results in a better taste is only true if coffee is roasted with this in mind. At which point I ask, is that really correlation anymore? It is important to note that I am using the Atago Refractometer and not the VST III and I also don’t filter my espresso before TDS measurements, so perhaps this correlation is true when it is filtered. But I really don’t want to filter coffee, its far too wasteful.

How precise do TDS measurements need to be? There seems to be a lot of discussion of needing to filter espresso to get precise measurements, however it isn’t clear to me how precise the actual espresso shots are suppose to be. Even looking at data I have seen online it doesn’t seem like there is the precision in shot TDS that there is in TDS measurements (unfiltered or otherwise).

When do you give up on a ML model? I don’t have a serious need to continue developing this model, so how would I know if the problem I am trying to solve is intractable?

For now I will just keep trucking.