Cooking With Data

As many of you already know, governance is always on my mind, and I find myself comparing governance to everyday life. I love creating analogies to help distill and simplify complex topics, so they are more digestible and understandable. And after a week in Stockholm, the correlation between cooking and the science of preparing, building and deploying AI responsibly became apparent. So here we are with a food analogy to help everyone understand the importance of governance in your AI adoption strategy. Now let’s first consider what preparing and cooking food truly entails. Cooking, also known as culinary arts, is the art, science and craft of preparing ingredients to make food more palatable, digestible, nutritious, or safe. Governing data is also an art and science and involves various skillsets. A chef might be cooking at a Michelin rated restaurant or an individual who simply loves to cook at home for their family. As technology has evolved, it has become easier for individuals at various skill levels to be involved in AI creation. GenAI is also bridging human agents into the loop to evolve and inform the model outputs. This increases the importance of an AI governance framework, just as cooks at various skill levels need guidance from a recipe to follow. This article will walk through the steps one might follow in cooking a meal, just as a scientist would follow a similar pattern in preparing, building and using the output of data from AI responsibly.

1. Evaluate the Recipe

Depending on the skill level, most cooks will follow a recipe to guide them. At times they might adjust the recipe to meet their specific needs, however, they still will need some level of instruction. When evaluating recipes, they consider factors like who are they cooking for? How many people? Are there dietary restrictions? When are they cooking this meal? As scientists are considering the use case for their AI, they will also have to evaluate factors like who is the intended audience? How will the output of the model be leveraged? Are their restrictions or considerations based on the audience, such as geography? It is essential the use case is fully understood, documented and transparent. Organizations need to have transparency into the purpose of these models to be able to properly understand and monitor the impact. Whether you are considering the meal you plan to cook or the model you want to build, evaluating these questions and having them well understood is an essential first step. For without knowing how you plan to use the model or which recipe you are going to cook, you cannot create a shopping list or have a comprehensive list of data you need.

2. Create the List ?

Once the cook has a clear understanding of the recipe, they need to document the ingredients they will assemble. This is a detailed part of the process to ensure no ingredient is forgotten. Imagine you are in the middle of cooking your meal, and as you reach for a critical part of the recipe, the ingredient isn’t there. Often, depending on how essential the ingredient was to the recipe, the entire meal could be ruined, and you must start over. Similarly, a scientist needs to know what data they need based on the use case on hand. Without a thorough understanding of the data required, data could be missing or there could be an element that is entirely forgotten. This is what can lead to bias in the system, because it is missing crucial data to feed the algorithm. Governance enforces thorough documentation of the required data that aligns with the use cases to help avoid this mistake. It is a framework to not only avoid this mistake but make gathering the data simpler task. How can you gather data, without knowing the data you need and how it will be used?

3. Gather the Ingredients

Now that the chef has documented the ingredients needed, they must acquire them. The ingredients come in various forms, often obtained from different stores. These ingredients are categorized into the departments, such as poultry, dairy, or produce. This makes gathering the ingredients easier and avoiding forgetting items on the list. Data is no different. Data will come in various forms in different locations, whether it is in the cloud, on prem, or in multiple data lakes. When the consumer is looking for the data to support their intended use case, they need a simple way to discover the data, regardless of where it lives. A governance catalog with logical groupings will help them browse, evaluate and access the data they need. Consider dairy as a Marketing Business Domain and cheese as the logical grouping of your Data Product. If you walked into a grocery store and all the ingredients were mixed with no system for categorizing, it would take a lot more time to gather. This system helps improve time to insights because scientists uncover the right data quickly.

4. Assess the Quality

As the ingredients are gathered, evaluating the quality of the ingredient is essential. For instance, what is the expiration date or is the produce too ripe or under ripened? Ideally the stores are only putting out quality ingredients that are not expired. There may be instances where product doesn’t require the perfect level of ripeness, so long as it isn’t under ripe. When a scientist is assessing data for their model, they will also be evaluating the quality of the data. In some instances, a quality score needs to hit a certain threshold, say 80% based on the use case. In other scenarios, a score above 60% could be sufficient. When data owners are publishing data into the catalog, it is essential the quality of the data is transparent so the consumer can evaluate based on their needs. Therefore, governance of data just as analyzing the quality of the food ingredients before selecting is important to a good meal or a valuable model outcome.

5. Prepare the Ingredients

A chef will need to prepare the ingredients before it is cooked. This could include measuring out the ingredients, dicing, mincing or staging them out based on what the recipe calls for. Preparing allows the cook to seamlessly follow the recipe and quickly identify the ingredients needed. Scientists also may need to prepare the data they have acquired before they build and train the model. This includes cleaning, transforming and integrating the data. This is a crucial step to ensure the right outcomes are achieved. Governance helps ensure that the data getting prepared has met the requirements for the use case that is thoroughly understood.

6. Cook the Ingredients

Now is an integration of the art and science of cooking. Based on the experience of the cook, they could follow the recipe to a T, or they can add a unique twist to the recipe. Regardless of the experience, the chef now has to take the prepared ingredients and cook them to the documented instructions. A scientist will now need to build and train the model accordingly. Human-in-the loop AI is an essential element to properly building models and avoid improper use. Just as a chef may need to let something simmer for 30 minutes, there is required observation to make sure it doesn’t over cook. The scientist is responsible for putting the same amount of rigor in the training of their AI. Governance empowers scientists to quickly get access to the right data and avoid missing data or using poor quality data.

7. Evaluate the Meal

Once the meal is cooked, the chef will then plate the food to present to the consumer. Yum! But wait! We cannot just assume the meal is ready. It is the responsibility of the chef, to evaluate the meal for quality, safety and taste. If poultry is undercooked, it could do harm to the consumer of the meal. At a restaurant, a chef needs to make sure the meal is prepared per the instructions, plated properly, and goes to the right individuals. Just as the meal needs to be safe to consume, the data must be as well. The data from a model is leveraged in various ways, and we must ensure that the model is being used as intended and guarantee the data coming out of the model meets the same bar as the data going into the model. With the thorough documentation of the use case, data required and process for building and training the model, any individual can evaluate the data output. After the output has been properly governed, the scientist can ensure it is safe for use or your meal is ready to enjoy!

Governance can be found at every stage of the AI lifecycle, just as can be found in cooking a delicious meal. Humans will come with various skill levels and experiences. This does not change the required steps; it just means the solutions need to meet the needs of diverse skillset. For instance, a data catalog needs to be easy to discover data regardless of their technical background. The use case must be easy enough to follow, evaluating quality and gathering data should be straightforward. Data is intended to be consumed, just like food. With the right framework and guide to follow, there should be no mishaps. Governance done right can ensure quality data in and out of the model based on intended use cases. Applying the same principles to food preparation will ensure an enjoyable, excellent quality and delicious meal! Okay, now I am very hungry! What should I make? ???