What is Explainable AI (XAI) and why does it matter?
Artificial intelligence involves computer systems that crunch through enormous sets of data, typically at high bandwidth rates, giving AI capabilities that you can consider to be superhuman.
For example, there’s no way a single person could analyze the thousands of purchases made in a supermarket chain’s California stores in real-time to detect patterns to squeeze more profits out of anything from produce to packaged snacks. But an AI model can certainly take on such a challenge. But how exactly does the AI reach its conclusions or provide a solution?
This level of power can be unnerving, especially as enterprises contemplate unleashing machine learning on ever larger troves of data to glean more value and insight from it.
When customers at large organizations contact us, sometimes I detect a note of worry in their voices as they ask about deploying machine learning for an artificial intelligence project.
They’re rightly concerned about the idea that using ML and AI can lead to developers creating a “black box” where ordinary mortals cannot comprehend what’s going on in the AI setup. When you ask data scientists to account for why an AI gave a particular solution or “output” to information it’s been fed, they are unable to give a good response.
It’s a weird business model to base decisions on conclusions where you lack true insight into the thought processes behind them.
Towards Data Science weighed in on the situation, noting that “As more and more companies embed AI and advanced analytics within a business process and automate decisions, there needs to have transparency into how these models make decisions grows larger and larger.”
At F33, we understand the issues that come when companies contend with artificial intelligence and a “black box,” since they’re not certain how it might affect their future growth and development, as well as their bottom line.
That’s why it is useful to become familiar with explainable artificial intelligence and what role it might play in your enterprise.
What’s Going on Under the Hood in Artificial Intelligence?
Non-programmers with a vested interest in how an artificial intelligence system works to analyze data and reach conclusions would be right to be worried about a program’s mysterious machinations.
As TechTalks put it, “Would you trust an artificial intelligence algorithm that works eerily well, making accurate decisions 99.9 percent of the time, but is a mysterious black box?”
That’s a decent success rate, except in cases when the safety of property and people is at risk. AI that prompts someone to make a financial transaction will not be so valuable if there is no context for why you should do one thing instead of another. An AI program used by a hospital to help guide the decision making process of nurses and doctors will be of no use if the medical professionals can’t understand what went into the thinking.
One potential solution involves a bit of a compromise. TechTalks reports that “Some research is focused on developing ‘interpretable’ AI models to replace current black boxes. These models make the logic behind their reasoning visible and transparent to developers.”
However, substituting an interpretable model for your existing yet enigmatic “black box” AI may lead to diminished accuracy. Therefore, organizations interested in AI will need to weigh the pros and cons of using explainable AI/interpretable AI and conventional AI on a case-by-case basis.
Principles of Explainable AI
Computer science professionals are working out the details of what explainable artificial intelligence should be. To that end, the National Institute of Standards and Technology has issued a report, which lays out four principles of explainable AI:
1. Explanation: The artificial intelligence system provides evidence and reasons for all the outputs it delivers.
An example of a suitable explanation would be an AI that suggested a movie for a streaming services customer to view. It would say that it made this recommendation based on criteria such as other movies you watched recently and what people in your area are watching this week.
2. Meaningful: An AI system offers an explanation that individual humans can understand.
In terms of being meaningful, the AI system would deliver an explanation that a person agrees another human would offer. If an audience comes to the same conclusion that the AI did, that lends further credence to its meaningfulness.
3. Explanation Accuracy: The explanation that the AI offers for the solution in its output is confirmed to truly reflect the processes used by the system.
The conclusions reached by an AI must be both high quality and coherent. Does the reasoning behind the explanation match with what a person would consider reasonable?
4. Knowledge Limits: This principle covers the idea that an AI system should only work under the precise conditions the programmers designed it to do.
The idea is that the AI should only give outputs when it calculates a threshold of confidence is met. Here, computer scientists are venturing into the field of metacognition, which means thinking about thinking, per the NIST. If the system for some reason functioned outside of its defined conditions, you can say that the results may not be reliable.
Deploying Explainable Artificial Intelligence In Your Enterprise
You wouldn’t dream of onboarding a new employee without looking at the potential recruit’s history via the CV and conducting interviews to see how his or her mind operates under different circumstances. For example: is this person creative under stress?
If you’re thinking of deploying a customized artificial intelligence system in your organization, it makes sense to have a similar level of transparency so you know what you’re getting into, just as you vet potential workers. To learn more about explainable artificial intelligence or to consult with us about AI and ML for use in your corporation, please get in touch with F33 today.