A match made in heaven: the marriage between machine learning and working memory
Let’s continue our discussion on cognitive computing from last week. Computers are significantly limited in their ability to learn abstract concepts — such as a big cat from just a few examples. Naturally, that limits the ability to understand marketing language without pre-training as there are more significant machine learning limitations when it comes to complex task solving.
Imagine this scenario: you wake up one morning and decide you want to see a big cat in real life — up close and personal. So, you look up the zoo on your phone and check out Google maps to find the best route. You catch the subway as instructed, get off at the station, and follow the street map to the zoo entrance. But a computer can’t do that. For a machine, learning the concept of a cat and a big cat is already a big deal, and so is learning how to read a subway map and learning to decipher a street map.
This complex scenario requires the equivalent of what us humans call a “working memory”. A “working memory” means we have the ability to store a certain amount of information, and we can then manipulate that information following certain rules.
The problem is that machine learning models in use these days simply do not resemble such a working memory — they are far from that. Remember, these models can at best distinguish a cat from a dog.
You can try to use them to solve complex tasks, but it would be the equivalent to a human trying to think with the part of the brain that is used to see and detect cats. In absence of a working memory, that part of the brain’s capacity to store information and the ability to manipulate that information is too limited to achieve reliable and accurate results that would bring you to the zoo.
What’s the solution to this problem? Cue the wedding bells. It’s the marriage between the concept of a working memory (something that has existed in IT for a long time) and the concepts of machine learning.
At Market Logic, our strategy is to help humans do the complex task solving and have the machine do the boring work: “see” concepts such as consumers, needs, brands and so on as the base line for further learning (brands, categories, barriers, drivers, and trends). This enables the machine to support the marketer with human level precision for concept understanding while teaching our machine to learn complex task solving (how to evaluate a new product idea, for example).
As a recent Harvard Business Review article notes, Market Logic and Unilever are effectively using this approach to augment human intelligence in insights engines to deliver real value for marketers.
Do you have any comments or ideas on cognitive computing? We’d love to open a dialogue.
Originally published on Market Logic blog.