Earlier we defined “creative” ideas as having novelty and value. Here, we explore overfit and how machines can generate creative ideas…and potentially take your job.
One really useful trick that higher order animals use to game randomness—to improve their chances of learning to survive and thrive—is called overfit.
We believe overfit is useful on an evolutionary scale. Organisms that wish to adapt to new environments may be presented with only a handful of life-threatening learning opportunities. If they do not infer conclusions quickly, without a large training set of data, they may die quickly. If they die before they reproduce, they are out of the gene pool. Life on Earth became capable of overfit a long time ago. Overfit’s side effect may be that organisms are likely to use aggressive survival tactics in many situations that are not life threatening, with the benefit that they will probably react to and survive true life threatening situations; therefore overfit will be selected for in the gene pool.
Think of overfit as an evolutionary reward for responding to logical false positive situations in order to always respond to actual threats!
True positive and true negative events occur when an observer’s prediction matches what actually happens. They probably result in an appropriate response to an actual event, and the observer has a higher likelihood of surviving that event.
False positive and false negative events occur when an observer’s prediction does not match what happens. They most likely result in an inappropriate response to an actual event.
A false positive was predicted to happen, but actually did not happen. “Nothing to see here, folks!” So the observer probably spent some energy on an unneeded response, but is still alive and well.
A false negative was predicted not to happen, but actually did happen. “Oh, crap…”. The observer probably ignored a situation that tried to kill them and, in all likelihood, suffered horribly and/or died.
In machine learning terms, overfit means that a created model may describe trends that do not exist in the actual system. Overfit may be the result of creating a model with training data that does not represent the entire target system, such as incomplete data or perhaps a biased selection of data.
There is also a concept of underfit in machine learning (and presumably in evolution), where a behavioral model is created that is too simple to capture the nuances of actual system behavior. The model is too general and too inflexible. Underfit can result in bad decisions, just as overfit can. But underfit models adjust to new data in an inflexible, slow-to-change manner. We speculate that evolution has not been kind to organisms that are not generalists and cannot adapt to new threats. Underfit most likely reduces false positives at the risk of missing an occasional true positive. Not good for long-term survival rates.
We believe that evolution favors overfit by the process of natural selection itself. Genetic mutation is random and only occasionally succeeds in producing a differentiated and viable new organism that is better adapted to its environment. Overfit may enable organisms to learn faster based on observations, and recent experiments suggest that generalized new learning may be passed genetically to offspring. We view this process of overfit and retained memories as an evolutionary hack to create new instincts in a rapidly changing environment.
Humans are perhaps the most adaptable organisms on planet Earth. We say “perhaps” only because we lack an objective scale with which to measure adaptability. We are human, after all, and biased slightly in favor of humans (though we try to be objective).
In our first post we talked about job specialization. In many respects, job specialization allows human organizations to behave like higher level organisms. People fill specialized roles in organizations, much like organisms have evolved specialized body parts and neurological functions. Some roles in organizations are tasked with protecting the organization and ensuring its survival, and some are tasked with experimenting to find new ways to differentiate and thrive.
The modern business environment offers a glimpse into how human organizations show overfit and underfit behavior:
Overfit: Risk-assumptive “one observation is data point, two observations are a trend, and three are a pattern”, where opposing parties all promote or defend their favorite viewpoints with the same data.
Underfit: Risk-averse “analysis paralysis”, where people cannot make decisions without more data, and when presented with more data never seem to have enough to decide. The usual decision here is to keep doing what the individual or organization has been doing, to “stay the course”.
We believe that overfit may play a huge role in enabling human creativity. Overfit may enable people to identify potential new patterns in order to learn fast. By doing so, overfit also enables humans to identify potentially inventive or innovative new ideas before there is enough hard evidence to support those ideas.
We believe that overfit may be at the root of successful “self-fulfilling prophecies” where an initial insight is followed through to its conclusion. The unsuccessful versions are called “delusions”, “daydreams”, etc.
The important part of overfit in human thought is that the correctness of an idea is not guaranteed or even desired on an evolutionary scale, as long as basic survival of a species is not at stake. People may want to feel like their ideas are correct, but often there is only circumstantial evidence that their ideas are loosely coupled to reality.
Overfit lets people build models from limited observations, and the testable models may produce unexpected new ideas and products that would not have been directly derivable even given larger and more complete observations.
On an evolutionary scale, humans are fast learners, indeed. However, human individuals show a lot of variation in risk assumption. While humans all must demonstrate some overfit in learning the basic functions and survival skills of life, some people like to base jump and others prefer a nice walk around the block. There is similar wide variance in high-level thinking and idea generation from person to person.
There is also substantial variation between humans in how quickly they draw conclusions from their observations, and how strongly they adhere to those conclusions given new and conflicting observations. Is flexibility or conviction a more valuable trait?
Once a human has created or adopted an overfit model for a given set of observations, they may tune the model based on additional observations. But they will rarely give up a model completely and start fresh. Creating an entirely new model is perhaps the definition of both existential and scientific revolutions.
How do employers assess employees for occupational fit with such a wide variation?
Occupational proficiency assessment is designed to test and assess what a person knows (knowledge and perhaps facts) and what they are able to do (skills). For most modern manufacturing and logistics economy jobs, this means measuring tightly focused “horizontal” skills.
Widely used assessment tests (educational, job placement, job performance) all measure innate sensory integration and learned horizontal skills (see our first post). They test what a person is physiologically capable of and how well a person has memorized knowledge and learned skills. Any test with a defined answer falls into this category, especially those tests that can already be graded by computer. We will dig deeper into assessment testing in later posts.
To summarize, we believe that…
Creativity is based on overfit, and
Overfit is machine learnable, so
People eventually will end up building creative machines.
Machine learning and robotics systems will become more capable of performing what humans consider to be novel and imaginative tasks. As this happens, job displacement due to automation will continue to increase. It is not a problem with human imagination. It is a side-effect of humans’ millennia-old refinement of specialized job roles that do not depend on imagination.