Dr. Babak Hodjat, a serial entrepreneur in Silicon Valley and chief architect of the technology behind Siri, believes that insights aren’t enough. In a past ODSC East keynote address, Hodjat leads us through what he believes the issues are and how companies can build true AI-Powered Organizations.
When you ask companies what they want from AI, many of their responses are prediction-based—know if a transaction is fraudulent or predict the risk of insuring a property. When they have these analytics, the company can make confident decisions on their own. They would take appropriate action to counteract the fraud or decide whether or not to insure that property.
The company doesn’t use AI to make the decisions themselves, only to provide the analytics to inform the decision. The trouble is that getting to those analytics requires a lot of effort—charts are confusing, graphs can be manipulated, and few people have the visualization training to read these complex outputs.
What’s the Problem With Analytics?
Most companies are focused on preliminary analytics, but this doesn’t do business impact justice.
- Most problems require multiple objectives—some kind of balance the company is striking between different solutions and needs.
- Models become obsolete faster than we expect—changes in customer habits, regulations, even unpredictable events like natural disasters alter the course of AI models.
Dr. Babak Hodjat believes that AI should be helping us make our decisions rather than serving purely an analytics function. He believes we’ve gotten some of this backward, and becoming a truly AI-powered organization is the way to fix it.
What’s the State of AI Now?
AI doesn’t entirely cover decision-making the way humans do. Three significant obstacles keep computers from making decisions on the human scale as well as we do:
- lots more trials required than humans need.
- not robust to unexpected states (i.e., states not in training models)
- not robust if the rules change (i.e., must learn everything from scratch each time.)
So how do we begin to apply AI to decisions when the state of AI hasn’t quite gotten there yet?
Understand Decision Making
Humans make decisions by taking context and applying strategies that optimize our objectives for making that decision. We create the plan through experience, but also mental models. We try out different choices within the virtual world of our minds and make the decisions. The outcome of our final decision is added to our historical data for future decisions.
We can build models that mimic this type of decision-making. In fact, this is the reality of data science! If we have sufficient data, we can create that kind of predictor.
So can we create a prescriptor as well? We must look at the ways humans strategize in creative ways to make our decisions.
Humans build decisions based on historical data. We tweak a small part of that history and see what happens. If it works, we keep it. If it doesn’t, we keep making changes until no better designs are found.
It’s a process known as hill-climbing, and this is currently the operation method for state-of-the-art AI. When you have simple problems, the hill-climbing method works well, but what happens with complex issues?
- The space is deceptive. AI doesn’t have enough resources.
- The space is too large. Too many variables create noise.
- The space is too high-dimensional. Single tweaks don’t yield enough improvement for the next course of action.
There are multiple objectives. The machine must delve into the art of possible.
Enter Evolutionary Computation
Evolutionary computation began at the dawn of computers with Turing. Instead of using one “climber,” it uses population search. Multiple variables are optimized at once, and multiple objectives and novelty can get around deception. You can remove the options that aren’t doing well, keeping the ones that are and allowing them to talk to each other.
It tackles an unfathomably large learning area, complementing deep learning to create these prescriptors. We can create multiple outcomes and run it against the predictor.
Evolutionary Surrogate Assisted Prescription helps machines make decisions more similarly to the way humans do through this type of computation. Data science creates a predictor model with historical decision making in the company. A prescriptor model runs against this predictor, creating predictions.
So Why Do We Need AI-driven Decision Making in AI-Powered Organizations?
Decisions are multi-objective, always. Dr. Babak Hodjatis clear that we must be able to use data to make the best possible decision. A lot is on the line for businesses to rely on intuition.
This is a way to do AI-driven models without waiting for inordinate amounts of data for training sets. When that data comes back, we can improve our models, allowing us to keep up with changes in the world that would make former AI models quickly obsolete.
AI-driven decisions help mitigate costs across business operations. We’re better able to make decisions without costly experiments in the real world.
Babak Hodjat asks:
- Where is the most impactful decision-making in your organization?
- What would that look like driven by AI-decision making?
Catch the video to see Dr. Babak Hodjat put this through a hypothetical situation similar to ones that a business might use (aka how an actual business might become an AI-powered organization) and join us for ODSC East 2020 for even more cutting edge AI design and thoughts on the future of data science in business from change-makers such as these.