Hacker News new | past | comments | ask | show | jobs | submit login
Better Beer Through GPUs: How GPUs and AI Help Brewers Improve (nvidia.com)
63 points by JasonCEC on Sept 2, 2015 | hide | past | favorite | 10 comments



Very cool! I thought it was interesting that you're getting good, usable data despite using untrained tasters - have you ever experimented with using some kind of calibrated tasters, like BJCP judges or something? Do you think it would lead you to get the same kind of data, but with a smaller sample size? Or would it not really make a difference?

EDIT: never mind, I just saw in another comment that you're using your own, trained staff - I miss-read the article.


We use our own staff for R&D (training the models), but semi-trained tasters are reviewing the product at the brewery;

we weight individuals on a verity of factors and have a pre-processing method that controls for experience level in expected flavor perception per flaw.

So experience level is important, but we often have to work around it. Most of the employees of a brewery or other beverage manufacturing are relatively experienced anyway!


Sounds a very interesting idea!

I'm curious what kind of volume of tasting data you're processing out of interest?

Also do you use textual reviews from beer rating sites too?

If the main idea is to detect flaws in beers, I'm wondering if you could also use mass spectrometer data, https://nextglass.co/science-of-satisfaction/ seem to be using them to compare the similarity of beers, although thinking about it, people's perceptions of different chemicals may vary widely.


We have a couple thousand reviews per style, so its really fat-data not big data. (in the tens of thousands for some styles)

The main difference is that we take a complete sensory profile from every review, and join that with environmental and personal effects, such as time of day, day of week, temperature, altitude, etc and learn the reviewers preferences and sensitivities over time. We take about 600 variables per review in total.

From that data, we're able to generate a much more actuate and actionable analysis (at a cheaper cost) than a GC/MS or HPLC could - just consider, your mouth is a perfect tool for tasting the compounds that matter to you when deciding what you like and dislike. This is modified by your environment, and we need to control for how your perception of flavor will change over time with age, exposure, and experience.

I believe that a GC is a waste of money for 90% of breweries.


Is "GC" short for gas chromatography?

If so, "MS" is probably "mass spectrometry" and "HPLC" means "high performance liquid chromatography".

Those were rather opaque to me, so I had to look them up, perhaps it saves someone the time.


Wow, 2 articles published about us in a day!

also being discussed here: https://news.ycombinator.com/item?id=10159852


How much of what you do is limited to being beer-centric? I run a small artisan cheese company, and having data like this would also be helpful at refining our make processes, as well as detecting seasonal preferences in both the milk we produce, and the cheeses that are derived from it.


Hey Sprocket!

We can do any homogeneous product made in batches - so cheese may be a good fit!

We've done some yogurt and cheese reviews in the past to test the system but nothing extensive.

Send me an email at JasonCEO [at] Gastrograph [dot] com and we'll see what we can do?

Cheers! - Jason


JasonCEC -- Just a question about the tasting protocols: where does the sampling and tasting take place? Is it local to the brewery and/or distribution areas?

Really fantastic idea and looks like a great implementation!


Hi gballard!

The tastings are done in-house by the companies production team. There are no samples to ship, and the sensory system is easy enough that just about anyone in the industry picks it up there first 4 - 8 times (its often more in-depth then they're used to!).

Thank you for the kind words!




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: