Machine Learning vs Learning Machines: The Power of Epistemology

Machine Learning is trending. I don’t do machine learning. I do machines that learn. There is a big difference which most of the media, investors, and data scientists don’t understand.

Machine Learning is pattern recognition, and it works about like a 2 year old does, or a being from another dimension that has no context. “People put food in their mouth, and poop comes out their butt,” is an observable pattern. It happens every time. Very consistent. Very predictable. Different foods come out at different speeds. Another pattern you can identify.

This isn’t particularly useful information. And while if you can pool enough if X then Y observations you might get to the point that you can identify that fiber and fat are two of the big factors in how fast things pass through people, you are never going to get to, “Why people choose to eat,” or “Why they choose what they do to eat,” and that is where the useful information comes from.

Context in humans comes from a hundred thousand places. Some are instincts some are shared preferences. Some are biological functions.

Consider an algorithm that determines how of the ultra-lush caviar ice cream you are going to sell at a given location on a given day. The location in San Jose California sells roughly 6 cones a day for each degree Celsius it is. The location in Manhattan does roughly the same, with some variance for which day of the week it is. When you ask the algorithm to pick where new locations should be set up it suggests Lake Havasu City in Arizona because it has the highest average temperature in the US. What Lake Havasu doesn’t have is $200k a year salaries who can afford the expensive ice cream, or the population to support it.

As a human we’d know this instantly. Just like we’d know that among the things we could do to increase sales, finding a place that is 100C is not among them since all the customers would die.

Deep Understanding, and learning machines do things differently. They have “checks” on their suggestions. “Will this kill people” goes on the list of things to consider, along with “is this legal,” and “is it cost effective.” A person trying to solve the problem of cellphones breaking from being dropped is going to rule out making the screens of diamond, and the body out of titanium, pretty fast because they have a number in their head for what the limit on the price to solve the problem is. Context means we self limit our choices.

The technical term is epistemology. An epistemology engine like I build has an understanding of what words and concepts mean. Maybe not to the level people do, but far more than what a dictionary does.

If I tell you that I want to take off work because the drummer from an 80’s heavy metal band died, you will ask if I knew him, because you know that isn’t a thing people would take off for grieving about normally.

If I tell you that I want to take off work because my mother died, you say, “of course, take as long as you need,” because you know that there is a close relationship between mother and child.

It may be “sad” that people die, but we don’t all take off work every time anyone dies because we’d never get anything done. That understanding of the closeness of relationships is just one thing that only people who experience some portion of it can relate to. Even understanding that if the person asking because their child died, and that would be more tragic than the loss of a parent, comes naturally to people, but machines don’t understand it unless they have context.

Epistemology engines combine pattern recognition with language processing. They use markers that are preprogrammed to learn new markers. For example, if the system notices an event in numerous people, “My mother passed away today” for example, and after that event their vocabulary and word choice changes from optimism to pessimism, and they speak less of social behaviors and happiness, the system learns that this is a “sad” event that is of high impact and a typical duration.

The system doesn’t know what it feels like to lose someone, but it knows how people respond. It is a mimic, not a true understanding. But as a result the system can look at how people respond, and know what its options are rather than suggesting inappropriate choices.

The system can also look at rules which were made by others to see what is going on. Having read the employee handbook for 50 of the top 500 companies the system can see that, personal days are granted for funeral attendance, and grieving for family members, but not drummers in bands who you don’t know.

The biggest difference in the two approaches, is probably that an Epistemology Engine works on unstructured data, which is what the majority of the web is, and Machine Learning works on structured data. As a result Epistemology Engines are not limited to answers dictated by the variables the data scientist gives the system, as a Machine Learning system would be, but rather can “think outside the box” and make determinations that are based on data the scientist didn’t consider.

If you want to learn more about my Cognitive Computing, Epistemology, and Artificial Intelligence, check out www.recognant.com

(reposted from https://www.linkedin.com/pulse/machine-learning-vs-machines-power-epistemology-brandon-wirtz )