Photo: Ariel Skelle/Getty Images
Eager to use artificial intelligence, some companies may be unnecessarily complicating easy business problems.
What's going on: Companies are over-using complex AI techniques when they would be better served with simpler approaches. Rule-based systems, for instance, show their work, thus allowing non-experts to pop the hood and see why an algorithm is misbehaving, unfair or biased.
Deep learning, the most advanced AI technique of the day, approximates how our brains work, with layers of "neurons" that together identify patterns. The catch is that their reasoning is opaque.
- That means a rudimentary deep-learning system for doling out loans could illegally deny applicants based on their race, just because it noticed that African-American borrowers default on loans more often than white borrowers. But it would be hard to know that was the reason behind the system's denial.
- By contrast, rule-based systems require an explicit set of instructions for making decisions ("if x is true, then do y"), making them easy to understand.
- Thus, while rule-based systems might also make flawed decisions, a user at least can devise them to deliver the best outcome — and then check precisely how it is doing.
Today, Adobe Research told Axios they now have a system that speeds up one technique for finding a good group of rules.
- The result: a 10,000-fold decrease in the time it takes to find a good decision set, allowing the use of far bigger databases, the way deep learning can mine enormous troves of information for patterns.
In a demonstration, Adobe data scientist Ritwik Sinha showed how the system derives a set of rules and uses it to divide up a jumble of data into sensible groupings.
- Eventually, a system like this might be used to help a company to divide up its customer base, understand why any one person belongs in a particular group, and target each with a different marketing campaign.
"There may be a problem for which accuracy is the only guiding principle," Sinha said. But when decisions directly impact human lives, an increase in transparency may be worth a decrease in accuracy.
What’s next: There will be growing focus on hybrid systems that integrate the power of deep learning with the simplicity and explainability of hard-and-fast rules.