A very interesting read, for those interested in applying artificial intelligence and deep learning to real world problems is “Prediction machines” by Ajay Agrawal. The book discusses the role that Toronto has played in the world of deep learning and how deep learning and AI has evolved and now can be applied to automate tasks at almost every level imaginable.
The last 50 years of computing has been mostly about converting everyday problems around us into arithmetic problems that CPUs can process much faster and more accurately than human beings. Examples are all around us: Payroll systems, library indexing systems, food store checkout systems are all handled largely by computing devices of several types. While somewhat automated, they have not achieved yet a degree of full automation as humans (and often human judgment) are still involved.
In the early days of artificial intelligence, the systems were primitive, and we followed a rules-based approach. This can have thought of as a series of If…then statements. If your car won’t start, then you should first check your gas gauge. If you have enough gas, then check your battery gauge, etc.…These techniques worked fine in somewhat finite universes of human judgment. A car has a limited number of reasons why it won’t start. But in a complex world, such as automating the driving of the car, rules-based systems are not adequate to the task as the possible permutations of human judgment required expand astronomically. Thus, the next generation of automation, which many engineers will take part, will be driven by viewing problems not as arithmetic in nature but more predictive in nature. At the core, what are you trying to predict?
Every morning, many of us try to predict the traffic patterns on our route to work. Others are trying to predict the prices of commodities or the company sales for the upcoming quarter or the resource demand in our retail store or the emergency room. All of these are notable examples of prediction models that could benefit from deep learning techniques. Deep learning can afford these systems the ability to adapt as more data is collected and to initially be coded to mirror human judgment based on a series of “training” procedures that are based on human judgment. If the system is properly designed, the system can learn from prior decisions and update its neural network.
The tool that was described by Dr Agrawal was the Deep learning canvas (https://docs.google.com/document/d/1yTwepxl8KDmv0QSTVorNWXrqaaW0Ryy8Dm7khLU93Ug/edit). This is a tool for working within organizations to help determine how to structure certain automation opportunities as deep learning applications. A simple example might be my grand-daughter’s grammar school. I noticed that security on weekends is almost non-existent, which is when the basketball games often happen. Imagine a camera and a procedure that could be fully implemented, requiring the person trying to enter the school to stand at a certain distance, remove their hat, and open up their outer coats and take their hands out of their pockets. The system here is trying to automate the determination of whether this person is a friend or foe (IFF). A bulge in a coat pocket, a clenched fist, an oversized trenchcoat could all be signs of a possible foe. The facial recognition aspect can match the face to hundreds of known or suspect criminals in a variety of databases. Compared to our current system of “neglect” which often occurs in budget constrained scenarios, this solution would require no human intervention until a “foe” is encountered. Thus, all the doors could be monitored and remotely controlled by a single individual who isn’t even onsite.
While this new world of automation may not be to everyone’s liking, the marketplace will be the ultimate arbiter of what gets deployed and what doesn’t. We will have to balance basic rights, privacy, personal safety and cost effectiveness in ways never dreamed of in the past. The realm of the engineer and engineering manager is to reframe the problems of the past into “prediction” frames. An example might be: Can I predict your next Amazon purchase? If so, why should I wait for you to order it and why not just send it to you on speculation? With a free return policy, we can start to achieve anticipatory selling. This is the ultimate in friction reduction as one click ordering becomes no click ordering.
My question to all those who have studied analytics or just have an interest in deep learning, what role will you play in changing our world? While the tools are powerful, the art will be in adapting the tools to the needs of society and will focus mostly on translating technical capabilities to real world problems.
About the author: Mark Werwath, Clinical associate professor, Director of Master of engineering management program, Co-Director of Farley center for entrepreneurship and innovation, McCormick school of engineering and applied science, Northwestern University