Show HN: Ml-hot-or-cold, Classify water as hot or cold from the sound of pouring | Hacker News Reader

Show HN: Ml-hot-or-cold, Classify water as hot or cold from the sound of pouring

github.com

7 pointsPNWChris6 comments

Hi HN, I'm sharing this project because it's finally in a state I'm happy to show to others. It's my first effort to make use of the skills you learn in Fast.ai's MOOC (part 1).

The Fast.ai course encourages experimentation to learn more effectively, and I have to admit, they have the right idea. I learned a ton working on this project that simply working through notebooks and taking the course couldn't make stick.

Feedback is very welcome. This is my first fully self-hosted web server, so I'm iffy on best practices for that part of the project. Plus, I'm sure there are plenty of domain experts on here to suggest improvements in other ways.

I've added some example data to the demo page, now folks can give the classifier a try without the hassle of making a file and uploading!

A direct link to the demo: https://ml-hot-or-cold.projects.chrisjeakle.com/

Was the same vessel used to pour the different waters? Same height?

It varied: in the “data/audio” folder I used a fairly heterogeneous set of vessels (a short thermos, short and wide ceramic Starbucks “you are here” mugs, some standard water glasses, and some somewhat narrow glass mugs). All were between 10 and 20 oz in capacity.

In “data-v2/audio”, I just did a ton of samples in ceramic Starbucks mugs and the narrower glass mugs.

It turns out the model generalized quite nicely! In build log 3 I took a model trained on “data-v2/audio” and applied it quite successfully to “data/audio”.

Is there any way to derive a human-comprehensible explanation of what the AI model is keying off of?

Alas, not in a way that I'm capable of, I'm still working through part 1 of Fast.ai's course. I looked over my folder of spectrograms, and there are definitely patterns, but I don't know exactly what features the model deemed relevant.

I've added some example audio to the demo page, so now you can see the accompanying spectrograms of various temperatures without needing to run the notebooks.

There is hope for human understandable models, however. Lots of brilliant minds are working on model interpretability [0][1]!

[0]: https://distill.pub/2018/building-blocks/ [1]: https://openai.com/blog/interpretable-machine-learning-throu...

Reply on news.ycombinator.com