Job Ads for AI Could Soon Look Like This. Are You Ready?
Wanted: Human Assistant to the Artificial Intelligence
We are seeking junior and mid-level human applicants to serve as data science assistants to our departmental artificial intelligence (AI) in charge of data analytics. Responsibilities include reviewing, interpreting, and providing feedback about analytics results to the AI, and writing summary reports of AI results for human communication. Requires ability to interact with vendors and information technology staff to provide hardware support for the AI. Experience collaborating with computer-based staff a plus. Must have good human-computer interaction skills. Formal training in the ethical treatment of computers and assessment of the fairness and bias of computer-generated results preferred.
The above is a job advertisement from the future – but not that far into it. It points to where we are going, and where we could be in maybe even as few as five years if we devote the resources and resolution to do the necessary research. But our recent past has shown us that we can develop the type of machines that would soon open up a whole new field of lucrative and fulfilling work.
See, over the last decade, a new computer science discipline called automated machine learning, or AutoML, has rapidly developed. AutoML grew organically in response to the many challenges of applying machine learning to the analysis of big data for the purpose of making predictions about health outcomes, economic trends, device failures, and any number of things in a wide field that are best served when rapid and comprehensive data can be analyzed.
For run-of-the-mill machine learning to work, an abundance of choices is required, ranging from the optimal method for the data being analyzed, and the parameters that should be chosen therein. For perspective, there are dozens of popular machine learning methods, each with thousands or millions of possible settings. Wading through these options can be daunting for new users and experts alike.
The promise of AutoML, then, is that the computer can find the optimal approach automatically, significantly lowering the barrier of entry.
So how do we get to AutoML and to the job advertisement above? There are several hurdles.
The first is persistence. An artificial intelligence (AI) for AutoML must be able to analyze data continuously and without interruption. This means the AutoML AI needs to live in a robust, redundant, and reliable computing environment. This can likely be accomplished using currently available cloud computing platforms. The key advance is modifying the software to be persistent.
The second hurdle is memory and learning. An AutoML AI must have a memory of all machine learning analyses it has run and learn from that experience. PennAI, which my colleagues and I developed, is an example of an open-source AutoML tool that has both, but there aren’t many others. An importance would be to give AutoML the ability to learn from failure. Its current tools all learn from successes, but humans learn more from failure than success. Building this ability into AutoML AI could be quite challenging but necessary.
The third hurdle is explainability. A strength of human-based data science is our ability to ask each other why. Why did you choose that algorithm? Why did you favor one result over another? Current AutoML tools do not yet allow the user to ask.
The final hurdle is human-computer interaction (HCI). What is the optimal way for a human to interact with AI doing data analytics? What is the best way for a human to give an AI feedback or provide it with knowledge? While we have made great progress in the general space of HCI, our knowledge of how to interact with AIs remains in its infancy.
It is entirely conceivable that an AI for AutoML could be built within the next few years that is persistent and can learn from experience, explain the decisions it makes as well as the results it generates, interact seamlessly with humans, and efficiently incorporate and use expert knowledge as it tries to solve a data science problem. These are all active areas of investigation and progress will depend mostly on a dedicated effort to bring these pieces together.
All that said, automated and persistent AI systems will find their place in the near future, once we make a concerted effort to thoroughly research it. We should start preparing our human-based workforce for this reality. We will need vocational programs to train humans how to interact with a persistent AI agent, in much the same way that we have programs to train others who work with and interpret specialized equipment, such as emergency room technicians. There will also need to be an educational culture shift on top of that training, as we will need to integrate AI interaction into courses covering communication, ethics, psychology, and sociology.
This technology is very much within reach. When we do reach it, we’ll have a new, expansive field for human workers. Soon, it will be time to write a job description, but only once we figure out some crucial problems.