Baking cookies shares a prominent similarity with training neural networks: Endless optimization potential. Neural network training is often not understood even by the initial programmers after the beginning stages.
Google has begun using neural networks to train new neural networks, resulting in lower costs of programming and faster time to functionality. Google has a black box neural network known as Vizier for this express purpose.
Vizier was used to train a neural network to optimize a recipe for chocolate chip cookies with a limited list of ingredients. Optimization problems tend to be computationally intensive, and so Vizier has been spread over several machines and a few different data centers.
In order to optimize the neural network, Vizier uses preset variables, known as hyperparameters, to find the best solution to a problem. These super-fine tuning adjustments can be made with respect to each component of a cookie.
Vizier adjusted ingredient levels such as butter, flour, and sugar, until a base recipe was found. After the base recipe was discovered, Vizier was used to pare down several million combinations into just 58 different batches of cookies over a period of about two months.
The challenge: Create a cookie with limited ingredients, including Cardamom and/or Szechuan pepper.
Vizier taught the pupil neural network how to optimize the recipes and helped to create a cookie that used non-traditional ingredients. According to Food Republic, the Google researchers involved remarked that the cookies were “Delicious.”
Which variables were tested?
Several factors were tested in this optimization, including, flour type, butter content, and chocolate content. The cookies had to comply with the bakery standards: be vegan-friendly, as well as, gluten- and soy-free.
Tapioca powder, flaxseed meal, white flour, and other forms of grains were tested to provide a sticky and well-textured cookie. The amount of butter was manipulated to prevent the cookies from being too runny or dry. Cardamom was added for a unique taste.
All of these variables resulted in a nearly-infinite number of possible combinations. Vizier helped narrow it down to just 58 recipes.
One other interesting ingredient was Szechuan pepper, also commonly known as Chinese coriander, which is added a bit of spice to the recipe. Szechuan pepper was not a final ingredient.
How can this apply to other areas of business?
According to IBM, optimization problem solving can benefit nearly any industry. Shipping can benefit from increase potential profit from optimally packed containers. Food service can benefit from optimal portioning and food storage, and marketing and advertising can benefit from optimized use of visual stimulation, phrasing, and more.
Cutting out unnecessary steps, reducing waste material, and increasing service efficiency are all potential benefits of using optimization in business. Less waste means a much higher return on investment as well as increased ability to explore new options.
Ultimately, Vizier made the recipe several million times, allowing it to sort through sometimes thousands of different options in parallel to find the best result. The remaining formulas were scored numerically based on taste and then refined into one, optimal, chocolate chip cookie.
Looking to make some cookies this holiday season? Why not use the Vizier recipe? Here it is, along with a link to the original release from Google.
The Pittsburgh-based bakery and cafe, Gluten Free Goat is vegan-friendly and soy-and-gluten-free.