Gartners’ “AI Hype Cycle” shows different AI technologies at varying stages on the curve. While AI Engineering is considered an ‘Innovation Trigger,’ Computer Vision is reaching the ‘Plateau of Productivity.’
As data leaders, irrespective of the vertical or size of our organizations, we are keen to explore opportunities to enable the ethical and responsible use of AI to support our businesses.
As we look to implement this cutting-edge technology and operationalize solutions, I would like to share some areas to watch out for. These range from operational misunderstandings to reputational concerns. I hope these will help you gain traction as you identify, implement, and scale your solutions.
In this article, I debunk three common misconceptions around AI showcasing how a Data Scientist is not the only one responsible for the model. It is a shared responsibility between Data Scientists and Domain Experts.
We often hear about “algorithmic bias.” I hold that this is a misnomer and should be struck from our vocabulary. And finally, AI is not a silver bullet for all complex problems in your industry!
Creation of AI solutions requires special skills and data scientists are critical! However, it is not just a technical endeavor. Instead, it is a collaborative effort between the technical teams and the business domain experts.
Irrespective of where Subject matter experts (SMEs) sit in the organization, they have to collaborate and be engaged with the scientists to ensure the desired outcomes are achieved and the algorithm is considering the nuances of the business.
For instance, when a data scientist is creating a model to optimize the supply chain, the business leader needs to help Identify the following:
Decision Variables: input variables for the model like equipment, materials, etc.
Desired Outcome: what is it that you want the algorithm to do e.g. maximize profit
Specify known Constraints: Labor restrictions, Equipment limitations, Geographical restrictions like weather, road conditions, or local laws
Data Sample: for training and validation of a model
The data scientist then trains the model and through that process, the model assigns a value for each decision variable given the constraints to maximize the business objective.
Without domain expertise, a data scientist can try different models but cannot be solely responsible for the validation of the outcome.
Beyond the outcome validation, as the model is tuned, the tolerance for loss (aka mistakes) is again something the domain expert helps with. For example, for a model that detects cancer in patients, a false positive is detecting cancer in patients that don’t have cancer and a false negative is the failure to detect cancer in patients that have it.
Are these equally bad? Is one mistake worse than another? Is the data scientist responsible for making this decision? This is where the misconception lies. It is the business leader’s job to decide the desired outcome.
Accordingly, it is the data scientist’s job to build and validate the model against the provided guidelines. The business leaders ensure the AI models are actionable and realistic!
We often hear about Algorithmic Bias and the need for “Responsible AI.” The second common misconception is the assumption that the algorithm is biased. Technology is not inherently biased.
Algorithms are mathematical models that learn from the data ingested based on the identified variables and specified constraints. The output of the model is a direct extension of how representative is the sample data you provide.
For example, while training a model to identify roses, if the training data is limited to images of red roses, the model will not classify a yellow rose as a rose!
Each model creation requires key decisions around variables to be used in the algorithm. In the simplified example of rose identification, I did not use color as an input variable. What about size?
If the training data used images of flowers in full bloom, the model would not be able to classify buds as roses. I hope you see how alongside the data, the Decision Variables play a role in the model outcomes!
Bias in AI is a subjectivity issue. It is a reflection of the conscious or unconscious bias of the domain expert in selecting data and defining variables. The charges of algorithmic bias that appear in the news are not inherent to the technology.
Rather, unexpected outcomes are most often the direct or indirect side-effect of decisions about data and variables.
The expectation from the Analytics function is the same, i.e., leveraging data to create insights for the enterprise that drive business outcomes. Thanks to the hype, there is an urge to look at every problem through the AI lens.
Data leaders and their teams lean towards AI to solve problems that can be solved by traditional methods. The basic approach to problem-solving is to use the simplest method possible. If a problem can be solved using traditional statistics models, SQL, data visualization, and analytics, we do not need to use AI, machine learning, or deep learning to resolve them.
Simpler methods have much better explainability and transparency. When your stakeholders come to you to understand the outcome, the deterministic nature of traditional methods allows you to sit in comfort as you elucidate the algorithm and walk back your output to the source!
As you move up the complexity chain of Artificial Intelligence -> Machine Learning -> Deep Learning, the accuracy might increase but explainability will plummet!
As Albert Einstein said, “If I had an hour to solve a problem I’d spend 55 minutes thinking about the problem and 5 minutes thinking about solutions.” This is sound advice!
My fellow data leaders, as you Marshal Forward in your quest to deliver value for your organizations, engage your domain experts and jointly validate the solution along with the input data, parameters, and decision variables.
It is a partnership with the domain experts, no matter where they sit in the organization that yields sustainable results for the organization. With every use case, assess the potential solutions and their efficacy and with the domain experts weighing the outcomes against the transparency of the solution.
About the Author:
A global technology leader, entrepreneur, mentor, and change-maker, Minoo Agarwal is passionate about all things Architecture and Data. With a proven track record in digitally transforming organizations building scalable architectures and using data-driven insights in the Healthcare-Life Sciences, Finance, Retail, and Marketing industries.
Agarwal has demonstrated repeated success in cultivating an outcome-driven culture for her teams in renowned Fortune 200 companies like Gartner, GE. and Labcorp. Agarwal is a published writer and respected speaker.
She has deepened her expertise to enable digital transformation via a University of Berkeley Executive Education certification, Chief Digital Officer Program; Future of Technology: Trends, Strategies and Innovation Opportunities; AI: Business Strategies and Applications.
She is a contributing member of the EDM Council. A Governing Body member of Evanta (A Gartner subsidiary) and other organizations.