Theory Thursday: The Business Value of Simple Models
Art by @basilonmypizza: https://lnkd.in/eF8FkWzN - https://basilhefti.ch/
CEOs ask me often: how does AI contribute value in my business?
When you are confronted with the daily announcements from frontier labs - better batteries, new materials, biomedical discoveries - you ask yourself: we do not have this R&D branch, we will never have such flashy cases.
This misses a key point: value is created in the detail.
Of course the grass looks greener on the other side. But time and again we realize: what we have is actually good. In fact, you may benefit from AI even when the final solution does not employ AI.
This sounds like a paradox.
A recent Nature Methods paper reports a related finding in biology, where the question was: “What happens to thousands of genes if we perturb gene A, gene B, or A+B?”
In the tested settings, current deep learning models did not outperform simple baselines.
This ties in with our own observations - and helps to answer the CEO’s question: AI can be very useful to find the essence of a problem, and to formulate a good method.
And the actual operationalized system may remain simple: rule-based, decision-tree-based, regression-based, or any other “boring” machine learning method - an advantage from an operations, maintenance and energy perspective.
So even when the grass looks greener on the other side: value comes from measurable impact, reliability, and cost-effective execution. No need to despair if flashy AI applications are seemingly absent.
As long as you get the business impact, you’re good.
In your quest for AI, are you looking for flashy AI cases, or for systems that improve decisions, cost, and execution?
See also: Trend Tuesday: Extracting the Essence of Software
• Art: https://lnkd.in/eF8FkWzN https://basilhefti.ch/
• https://lnkd.in/eAEPx9Fk, C. Ahlmann-Eltze, W. Huber, S. Anders, Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines
