s22-team8-project

Team #8 - Inverse Nutrition Mapping

This neural network is designed to take in information from nutritional facts labels, such as the nutrient information and ingredient order, and the network will return the mass in grams of each ingredient. Specifically, this network was limited to focus on breads, so the selection of ingredients was limited to the fifteen ingredients most commonly found in bread. Similarly, the nutrient information was limited to only include nutrients that the Food and Drug Administration requires to be present on all nutritional facts labels.

The following table represents the ingredients and their respective nutrients:

Ingredient	Calories	Total Fat	Saturated Fat	Cholesterol	Sodium	Total Carbs	Fiber	Sugar	Protein	Vitamin D	Calcium	Iron	Potassium
Active Dry Yeast	295	4.6	0.6	0	0.05	38.2	21	0	38.3	0	0.064	0.017	2
Water	0	0	0	0	0.003	0	0	0	0	0	0.003	0	0.001
Salt	0	0	0	0	38.758	0	0	0	0	0	0.024	0	0.008
Bread Flour	361	1.7	0.2	0	0.002	72.5	2.4	0.3	12	0	0.015	0.004	0.1
Butter	717	81.1	51.4	0.215	0.576	0.1	0	0.1	0.9	0.000056	0.024	0	0.024
Sugar	375	0	0	0	0	100	0	100	0	0	0	0	0
Egg	143	9.9	3.1	0.372	0.14	0.8	0	0.8	12.6	0.000035	0.053	0.002	0.134
Dry Milk	496	26.7	16.7	0.097	0.371	38.4	0	38.4	26.3	0.000312	0.912	0	1.33
Wheat Flour	364	1	0.2	0	0.002	76.3	2.7	0.3	10.3	0	0.015	0.005	0.107
Vegetable Oil	884	100	7.4	0	0	0	0	0	0	0	0	0	0
Olive Oil	800	93.3	13.3	0	0	0	0	0	0	0	0	0	0
Honey	304	0	0	0	0.004	82.4	0.2	82.1	0.3	0	0.006	0	0.052
Brown Sugar	380	0	0	0	0.028	98.1	0	97	0.1	0	0.083	0.001	0.133
All-Purpose Flour	364	1	0.2	0	0.002	76.3	2.7	0.3	10.3	0	0.015	0.005	0.107
Milk	50	2.1	1.2	0.008	0.047	4.9	0	4.5	3.3	0.000001	0.119	0	0.057

A typical input is composed of two tensors. One represents the nutritional information and the other contains the ordinal ingredient information.

The nutritional information refers to the quantities of each micro- and macronutrient whose disclosure on nutritional facts labels is mandated by the FDA. An example of what the input for this tensor might look like appears in the table below. In that example, the recipe has 500 Calories, eight grams of total fat, and so on:

Calories	Total Fat	Saturated Fat	Cholesterol	Sodium	Total Carbs	Fiber	Sugar	Protein	Vitamin D	Calcium	Iron	Potassium
500	8	4.5	0	0.05	30	18	2	38.3	0	0.1	0.02	2

The ordinal ingredient information simply represents the relative magnitude of each ingredient. Each of the fifteen ingredients has its own index in this tensor. The element's value is set to zero for any ingredients not present in the recipe. Then, ingredients are assigned integer values in ascending order, so that the ingredient(s) with the smallest non-zero mass in the recipe are given a value of one, the next largest mass is given a value of 2, and so on.

The output of the model will contain information structured like this, with the masses of each ingredient given in grams:

Bread flour	Water	Butter	Sugar	Salt	Dry milk	Active dry yeast	Egg	Whole wheat flour	Vegetable oil	Olive oil	Honey	Brown sugar	All-purpose flour	Milk
408	236	14	12.6	8.6	8.48	5.6	0	0	0	0	0	0	0	0

Running the Demo

1. Python 3.0 or higher is needed. This can be found on the Python website
2. The project is written in .ipynb files that require Jupyter Notebook, which can be installed from here
3. Once the dependencies are installed, navigate to and open the "Project Demo.ipynb" file in the "Project_Demo_Folder" directory and run each cell by clicking on the ▶ button above the code/text.

Alternatively, the demo can be viewed (but not run) from GitHub through the Project Demo notebook

Companion Files

The copy of the nutritional information table given above that is used by the project is saved in the nutrients.csv file.

Testing data was created by generating random bread-like recipes. These recipes all contained yeast, a liquid, and a flour, and optionally additionally contained a sweetener, lipid, salt, egg, and dry milk. A small number of real recipes were used for validation as well.