Baking with machine learning (2020)

> I know, I’m using cups and teaspoons and not something more universal like grams. If my little model becomes a smash hit, I will try to add metric system support.

It's not that cups/teaspoons aren't universal that makes them noteworthy, it's that they're volumetric. You should be measuring your dry ingredients by weight. The unit isn't that important, we can convert once you start measuring properly.

Fun fact, measure of a teaspoon is confusing:

1 US tablespoon = 14.8 mL = 3 teaspoon

A metric tablespoon is exactly equal to 15 mL

The Australian definition: 1 tablespoon = 20 mL = 4 teaspoons

> https://en.wikipedia.org/wiki/Teaspoon > https://en.wikipedia.org/wiki/Tablespoon

Now my so gave me a weird look. I had all the ingredients in the bowl and I said, "It's a cookie!"

An interesting normalization of amounts for baking bread and similar foods is called “baker’s percentage”.

The total weight (or mass) of flour in the recipe is 100%, and all ingredients are measured in terms of that. You do end up with very small percentages for ingredients like yeast, but in general it’s nice because you can easily see for instance how sweet a recipe is based on how much sugar is added as a percentage of the total flour.

https://en.m.wikipedia.org/wiki/Baker_percentage

Seems like a linear model would work equally as well, without all the deep learning bits.

There's a more practical but relevant cooking problem that I haven't seen discussed - when extracting recipes from websites it's tricky to parse the actual ingredients and quantities.

For example, "1/2 cup of diced tomatoes or a can of tomatoes (preferably san marzano)". This sort of freeform text doesn't suit regex very well but also lacks substantial context clues. You'd most likely use named entity recognition which could recognize that "1/2" is a quantity, "cup" is the unit, etc. but I haven't gotten very good results yet.

Maybe I'll write up a post when I land on a solution.

The next logical step here, it seems to me, is to scrape tens of thousands of recipes and apply this more broadly. There are, alas, perhaps a dozen or more dimensions along which recipes are going to be difficult to sanitize and normalize—there’s weight vs volume, variations in how to describe similar ingredients, and perhaps worst of all is the actual steps!

But are these challenges insurmountable? Well, maybe a few would limit you in some respects, but it certainly can’t be denied that some intrepid coder-cook could push this concept somewhat further.

You could also turn this all around and have a “consensus recipe generator.” Type in “pesto genovese” and it gives you a best guess of ingredients and amounts, with links to some recipes roughly matching those consensus amounts, perhaps with notable variations/“nearby” recipe clusters (sun-dried tomato pesto? or just the most popular pine nut substitute?) noted too.

I noticed something about Adam. Suppose you want to do gradient accumulation. Easy: compute the gradients of the loss with respect to each model parameter, and accumulate it over N training examples. Then pass the result into Adam, performing a single step. This is standard gradient accumulation; it's equivalent to running a mega-GPU that can process 64 training examples at once, rather than a small-GPU that can only process 8. It just takes longer to train, since you're performing an adam update every 8 steps in that case.

But, I was staring at the Adam formula and thought of something. I'm not sure if it makes sense, but there seems to be an "alternate" way to accumulate gradients:

For each training example, compute the gradients and apply the gradients. However, we apply them in a special way. The final step of adam normally looks like this:

  param = param - (lr * m_t / (v_sqrt + epsilon_t))

I propose accumulating the gradients like this:

  accum = accum + (lr * m_t / (v_sqrt + epsilon_t))

Then after N training samples, when you want to do the actual variable update:

  param = param - accum
  accum = 0

The advantage of this approach (if it works at all) is that Adam updates continuously. Every training example would cause Adam's mean and variance estimators to update. (Recall that the whole point of Adam is that it tracks mean and variance for every parameter in the entire model.)

So, in traditional gradient accumulation, those mean and variance slots would only update every N training examples. With this approach, they would update every training example, and then the model params update every N training examples.

It might seem like a small tweak, but adam's variance stats are crucial; it's what makes adam effective. Updating the variance 8x more frequently might be an advantage.

I recently used AI Dungeon to obtain an omelette recipe. You follow the resulting instructions at https://play.aidungeon.io/main/adventureView?publicId=6a346e... at your own risk.

Side note:

one project I would like to do myself is a followup to YY Ahn et al's Flavor Networks research, but this time with some extra.. `zest`: using the flavor and ingredient networks, but turbocharging the model using some kind of molecular representation of the flavor compounds using e.g. dgl-lifesci for graph-based representations of individual molecules.

Basically instead of

food <-> ingredients <-> flavor molecules (strings) relations

use

food <-> ingredients <-> flavor molecule (graphs)

since graph neural networks are all the rage these days. I already made a kind of graph neural network thingy that works similarly for the AICrowd Learning to Smell competition that was kind of similar :) https://www.aicrowd.com/challenges/learning-to-smell

In case you're wondering why this is circulating, Google had it on their cloud blog a couple weeks ago:

https://blog.google/products/google-cloud/just-desserts-baki...

I wrote it up around then too. Haven't tried the recipe, but the cakies look legit.

An interesting application is the target of a feature set configuration that returns a specific desired output activation value (eg 50/50 for two of classes). I do not know if tensorflow has an automation for this task available, perhaps could be a fun followup project for the author.

Looks like the article is almost a year old, we should add "(2020)" to the title.

I was hoping for a home made robot that cooks at home...are we still that far from it?

I honestly read banking....

I guess this is not a reference to the TDP of some video cards.

go sara!!

Well we know what happens next. Some ambitious kid, in some random corner of the world, is going to cut and paste it and sell it to some police dept as a "Google" profiler. It will work as badly as the police do, but now they just point at the app to cover their ass, cause you know Google said so Judge. Where it doesnt do what the cops want the kid will add "features" to make sure it does. Give it a year or two and the local PD will be happily be promoting it to everyone else for a decent commision ofcourse. "Whistleblowers" will get taken care off. VCs will step in to scale up ops. The price will be so cheap it will bankrupt most honest competition. Half the world will be using it soon enough and then ofcourse Palantir will step in and buy them. The billionaire founder will then strut around grooming the next gen, funding further mindlessness. And the same thing will repeat from Medicine to the Military, from Finance to Psychology all the while a clueless buffoon class of corporate programmed robots stand up and talk about how they are empowering the masses.

So how do we stop the cycle? Dont empower the masses. Just yet. Dont work for execs pushing such narratives. If you are in positions of authority sideline them. They dont know what they are doing. Scaling everything is a prime directive for these robots and they are too dumb and unimaginative to rewrite their own code let alone decide what is good for the population.