The explosion of artificial intelligence (AI) initiatives at enterprise companies has brought a wave of new opportunity for data scientists, as well as a lot of frustration. Strategic data science projects are exploratory, high-visibility, and high risk — and it’s unlikely that they will end with a perfect solution. Data science teams want to accelerate delivery and reduce risk of these projects, but the traditional approach to AI development is hindering their success.
Data Scientists’ Frustration Identified
The common approach to AI project development is linear and sequential. Also known as “waterfall development,” it is long familiar to software developers and now often applied to AI development. This approach flows in one direction, like a waterfall, through each phase of the AI development cycle, and requires an enormous amount of initial planning. The entire solution must be complete before it can be launched and deliver value, which means defining all models and anticipating all issues upfront. The process then flows downstream from stage to stage, team to team, until all design, development, and testing phases are complete.
Not only is this method rigid and restrictive, it can take many months before enterprises see value. They also risk building the wrong model. The waterfall approach requires teams to predict major obstacles rather than taking a step-by-step, iterative approach that enables the discovery of problems early and allows for course correction along the way.
Agile AI Approach
Over the last decade, software developers have felt similar challenges and shifted their process to "agile development." This approach breaks down the larger effort into small, bite-sized components that are finished and quickly launched, providing teams with flexibility and the ability to ship software on time. Additionally, this allows for subsequent launches to continuously improve as new components are added.
Successful AI teams observed the success of agile development and began to apply the same principles to their projects. The “agile AI” approach allows data science teams to continuously learn while giving them the opportunity to re-prioritize where necessary, and ultimately deliver value faster and more efficiently by building the right solution.
How the Agile AI Approach Works
As an example of how an agile AI project may work, imagine a company that needs to review hundreds of thousands of user-generated photos before featuring them on a website or social channel. This might require checking for potential violations including explicit material, copyrighted images, or product or brands that could cause conflict with its advertising partners, etc.
This internal image review process is time-consuming, expensive and lacks scale as volume increases. To make it more efficient, the data science team automates the process by establishing an AI model. Under the traditional waterfall approach, the team designs their AI solution with a comprehensive model to address the three core criteria under review for each image. To train the model, the team must generate a large amount of example data for each violation type and the various combinations. Once the appropriate training data is incorporated in the model design, the testing process can begin with model scoring and validation, accounting for variations of all three violation types so the model can be improved. Now the solution is ready to be deployed and can begin to add value.
With the agile AI approach, teams begin by identifying and prioritizing the distinct violation types that need automation. For instance, the business team could decide to prioritize identifying advertising conflicts, since those have a direct impact on revenue and are very time-consuming for human moderators to pinpoint. The data science team would then develop an AI model using only training data for advertising violations while an internal quality control team handled the other use cases.
In this scenario, the data science team will generate one training data set for advertising imagery, allowing them to test, score, and validate a simpler model. This first model deployment will provide a suggestion engine that attempts to determine advertising conflicts, while the human quality control teams confirm or correct manually as they review each image. This method reduces deployment time and increases efficiency — and the AI model continuously improves with human feedback. Eventually, the roles are reversed, with humans only reviewing the exception and low-confidence results.
As this model matures, the data science team will then focus on data for the next violation type via the same iterative and incremental approach. This course of action is repeated until the entire solution is complete — or updated — based on business value and real-world observations. By breaking down the AI solution into prioritized components of incremental value and repeating the three AI stages of train, test, and exception-handling for each, delivery of business value is accelerated and risk is reduced.
Although the agile approach is common for software development, it’s still rare in AI development. Data science teams that adopt the agile AI approach are deploying more accurate and valuable solutions faster and allowing teams to more rapidly integrate AI into other business functions.