Keeping Humans in the Loop

One way to avoid tragic or just frustrating AI fails is to keep a human in the loop for corner cases

To unlock the real power of artificial intelligence, especially in the near-term, humans will have to work closely with AI. There are many specific tasks where AI can perform them better, faster or cheaper than humans. However, these specific tasks are usually part of a larger function done by humans. To be effective, AI usually requires that certain conditions be consistently maintained, for example lighting for computer vision or context for a conversation. For most applications, AI has difficulty in corner cases and a human is needed to step in on occasion. Also, sometimes the implications of failure are simply too high to completely abdicate a task to AI, for example medical diagnosis.

Interface between human and AI may be more important than the smarts in the box

People have many qualities and capabilities that are still very hard for AI to imitate. Chess Grandmaster Garry Kasparov, who lost to Big Blue the IBM AI system in 1996, has argued that the best chess players will be teams of humans and artificial intelligence. This partnership is called augmented intelligence. When designing an AI solution, it’s best to make one that works well with humans, and gets the best out of both humans and machines. In other words, the interface between human and artificial intelligence is critical. We need to spend significant time on this interface, perhaps as much as the internal algorithms and performance capabilities of the artificial intelligence itself.

Sawyer, an early version of a robot capable of working alongside people (aka cobots) developed by Rethink Robotics (since acquired by HAHN Group). A user can grab the arm of the cobot to demonstrate tasks using a very intuitive interface. The task could be as simple as picking up objects from a moving conveyor belt.

Though it's unclear whether Sawyer will be a commercial success, it is an attempt to migrate AI and robotics toward more generalized tools that can be used by ordinary, non-technical people to augment their work. In the same way the spreadsheet did not replace accountants, cobots, whether physical machines or just software, may free up people to do more interesting and valuable work.

When failure is not an option, as in healthcare, then best to augment human expertise and judgment, rather than replace it

One area where augmented intelligence is particularly relevant is healthcare. Understandably, people are reluctant to have bots diagnose or treat diseases without human doctor oversight. At the same time, hundreds of thousands die every year from medical mistakes. Much of the work of doctors and other caregivers has been systematized for consistency and cost management, but this may make the work dull and monotonous. Great for the science and business of healthcare, but not so fulfilling for caregivers or ultimately good for their patients when fatigue or boredom leads to mistakes. This is an area ripe for AI. Let the bot do the systemetized portions of healthcare, and let caregivers do what they do best – using their expert judgment and providing empathy and care to their patients. This should lead to better health outcomes while making the caregivers’ jobs more enjoyable.

One example of augmented intelligence in healthcare is a company in which I have invested called HealthTensor (since renamed Regard). The company is focused on improving the hospital doctor's regular check-ins with patients. This is a relatively controlled situation, rich in patient data and already systematized with clinically-validated protocols. These doctor check-ins typically require that the doctor review the updated vital signs and other data on the patient, and based on the protocol decide whether a change of medication or other action is required. The doctor is then required to document the review and any decisions.

HealthTensor streamlines the doctor’s workflow by collecting the patient’s data, applying clinically-validated algorithms to diagnose and make recommendations, which are then reviewed by the doctor for modification and signoff, all the while documenting the visit for the doctor. The impact of this system is that it augments the doctor’s expertise and judgement, but does not replace it. It reduces error and improves consistency for macro-studies and improvement of protocols. It may even reduce costs by allowing doctors to care for more patients and potentially liability for malpractice. But most importantly, it may save lives.

AI can learn from interacting with humans, especially in retail where intimate knowledge of products and being biological helps in the long tail search

Retail is another area where augmented intelligence may prove very relevant. A big part of retail is the process of helping consumers find products that fit their needs. This search process often takes the form of a conversation between consumers and retailers exploring the many attributes and applications of products. Consumers may have very specific needs or desires and a good retailer needs to understand their products very well to help in the search process. Much of this information is informal and locked in the brains of retailers. This problem is often referred to as long tail search.

One startup based in Toronto called Adeptmind, is seeking to unlock this information using AI in tandem with retailers. Using AI alone for long tail search is problematic, since there is little data given how few of these complicated searches are repeated. Adeptmind solves this problem by having a human in the loop, in this case retail staff who know their products well. In a service they call guided discovery, the retail staff can be part of the process in real time, while the AI makes suggestions while learning at the same time. It's a bit like online chat enabled by AI. Overtime, the AI will be able to answer more search queries without any human intervention.