Machine Learning for Blind Tasters

Question: What do you need to improve on in tasting?

In my personal blind tasting progression, and in running a weekly blinds group with both MS and MW candidates, I've often made comparisons to machine learning. This analogy is incredibly useful for figuring out how to get better. 

[Note: Geoff Kruth in particular espouses an extremely logical approach that I really identified with, and a lot of what I'll lay out here is based directly or indirectly on his perspective.]

Machine Learning in brief:

We'll focus on one type of ML model: A classification model, or classifier. A classifier uses processed data, or features, to decide which of several classes the input belongs to. 

I'd encourage everyone to walk through this excellent visual introduction to ML: http://www.r2d3.us/visual-intro-to-machine-learning-part-1/ which explains the fundamental concept of features, decision trees, and overfitting. It's better on desktop than on mobile.

Deductive tasting as a classification problem

The analogue for blind tasting is:

  • Raw data = sensory perception (e.g. green bell peppers)
  • Features = objective factors (e.g. pyrazines)
  • Classes = styles (e.g. NZ Sauv Blanc)

Note that this is more than a loose analogy - I've actually built this model.

Results (93% accurate):

What this means:

We can apply some intuition from machine learning to blind tasting:

1. Separate Sensory from Logic

Don't second-guess your sensory based on your Logic, and vice versa. Assume your sensory is correct and make the best logical call you can. Then afterwards, identify whether you need to improve Sensory or Logic. 

2. To improve results, Features > Algorithms

For advanced tasters, perfect logic should be a given. If your sensory is good, you should be able to identify classic styles almost every time, even accounting for some variation in structure and aromatics. To progress, focus on improving your ability to identify objective factors, which involves improving identification and calibration. 

3. For training data, Quality > Quantity

Taste representative wines.

4. Beware of Overfitting

Not everything is important. Avoid changing your logic just because you got a single wine wrong. 

 

There are lots of finer points that fall out of this analogy, but I'll stop here for now. Returning to original question, what is your personal assessment of what you need to improve on in blind tasting?