How to Accelerate AI Model Development Cycle

AI assets are essential for providing competitive advantage. Building quality AI assets quickly becomes crucial in generating this advantage. While there are many tools available to do so, we’ve seen the success of the following four approaches and we believe they are different.  

The first approach to accelerate the AI modeling approach is to focus on problem solving other than model building. The most common mistake slowing the AI modeling process is that the development focuses too much on fitting a model rather than solving a business problem. This is usually a consequence of lack of communication between data scientists and the business owners, or of the little time spent in business understanding and framing a problem to create incremental value. While it is sometimes easy to define and develop a machine learning model, it’s much harder to model the right business question so that it can directly be incorporated into the business process and can add value right away. For example, many companies try to forecast sales. While it’s relatively straightforward to set up a machine learning model to forecast sales, it’s very hard to leverage the model to create additional value as it does not necessarily model specific decisions one could make. Instead, we’ve seen much greater success when you directly model the outcome (ex. incremental sales) of a specific decision (promotional activities).

The second approach to accelerate the AI modeling approach is to spend time developing AI foundations that can help data scientists to focus on solving the problem. It is widely known that the majority of the time spent by a data scientist is not on modeling or I call solving the problem. Rather, it is spent on data preparation, data cleaning, back and forth data understanding, data engineering, etc. While it’s very important to improve data literacy, it is equally important to provide the right tools to the data scientists so that they can focus on modeling. When designing these tools, we need to prioritize any repeated steps data scientists need to make (feature engineering, model automation, model explainability, etc.). The ideal solution is a customized and robust system that is aligned with your internal data management solution, computing solution and BI solution. 

The third approach to accelerate AI modeling development is to decompose the big problem into smaller problems. Three goals involved in this process make this decomposition more challenging than it may sound. The first one is feasibility. It’s usually very important to have a proof of concept (POC) as quickly as possible before a lot is invested. Decomposing a problem so that the first chunk is doable and feasible to finish in a couple of months is important. Some data scientist teams develop 1-2 models in a year based on my experience, which is too slow from a POC perspective. At the same time, each decomposed step should be feasible and can be achieved in a couple of months. Second is expandability. Building a POC is not the goal but a first step to achieve a bigger goal. Analytically this means the model build has to be simple but can be easily expanded so that once POC is successful there is an easy route to expand the model to solve additional questions or add additional features. Third is value generation. Each decomposed model should generate incremental value for the business and the decomposition should each be incremental to the previous step. This requires very careful design based on specific business problems that’s been solved. Feasibility, expandability and value generation together define how the problem should be decomposed. 

The fourth approach to accelerate AI modeling development is to redefine the coding philosophy. AI coding is different from software engineering. Data scientists don’t necessarily know what results to expect and typically need to try a lot of analytical assumptions when solving a problem. Software engineering is, however, focused on engineering to achieve specific functions and typically very little science is involved. AI coding philosophy is that many assumptions can be tested in a short period of time and that when tests are finished the code is scalable and we can move forward to the next step. This philosophy means that data scientists should write concise but scalable code as some problems require a lot of assumption testing and learning. Many machine learning tools provide scalability but they also sacrifice capabilities to try many different ideas by restricting the coding structures or sometimes models to choose from as well.