When it comes to learning and understanding the world and its concepts, humans have a fascinating capability: we can learn from just one or two examples. This enables us to be successful, even in uncertain conditions. A simple example is that we are able to identify a domestic cat after being given only few examples to learn from — our parents point at a cat in the street and say “this is a cat”, and then do the same with a picture of a cat in a magazine. Even 4-year olds will then, after only two samples, have developed (learned) an understanding of the “concept” cat, and that’s quite helpful of course, because sometimes, we only get limited chances to understand the world around us. Think about deciding whether an animal is a cat or a tiger … you better make the right choice right away!
But this fascinating human capability is an extremely challenging problem in cognitive computing. These days, most machine learning algorithms need to ingest huge amounts of data about isolated concepts and then connect the dots. For our cat story, you’d have to show the algorithm hundreds of thousands of examples of a cat. Gradually and with extremely intense computation, the machine will learn to identify the unique features of cats on a more abstract level, adapt the decision criteria over several learning iterations, and then store those weighted rules, so when you show it a cat it hasn’t seen before, it will still be able to identify it as a cat. Unfortunately, when you show it a picture of a tiger, chances are it will classify this as a cat as well…
So, how do humans do that differently?
Well, humans, after a certain amount of time in this world, stop learning from scratch every time, while computers always do. Prior to seeing the cat, our 4-year old has maybe learned what a dog is and can abstract the learning to a meta-level: “How is cat different from dog?” That is a much easier task, and over time, the speed of learning new concepts reaches impressive levels, because the more abstractions you learn, the quicker you can spot the new features in a previously unseen concept.
This is also the human capability that enables a marketer to learn the difference between the concepts of a “problem” and a “benefit” for a product. At Market Logic, we are making our systems capable of real one shot learning. Additionally, we train the machine as a managed service so that understanding of marketing is built right into the knowledge graph we call the Market Logic. It’s also a key reason why Harvard Business Review says our insights engines deliver real value for clients like Unilever from day one.
How do you use cognitive computing in your everyday work life? We’re interested to hear.
Originally published in Market Logic blog.