Evolution, Machine Learning, and Us
One of the benefits of being interested in lots of different things, or simply paying attention to topics in school even when they aren’t your specialization, is that you’ll start to notice interesting connections between seemingly unrelated things.
Something that I’ve noticed is that the process of evolution, how people learn and develop, and machine learning all share common features.
Beyond simply the mechanisms in which they work are also the relationships between how they came to be. Machines were designed by humans. Humans, in a sense, were designed by evolution. The term “neural network” originates from human minds in both senses of the phrase: humans intentionally studied the neurons of the brain and borrowed its structure for neural networks. Between computers and evolution directly, there is even a branch of study called “evolutionary computation” where computer models simulate evolution.
Let’s go over a simplified version of each of these processes to see how they are similar. Below I’ve split up the content into where processes start out: the beginning, when life first begins, when a human life first begins, and when a piece of code has yet to begin training; the testing phase, where learning and trial and error occurs; and finally, the iteration section which makes updates to the previous form and improves the form upon each cycle.
To be clear, in the following sections, 1, 2, and 3 correspond to evolution, humans, and machine learning respectively.
The Beginning
1a: Evolution begins with life. Life is really simple at first, just these little containers of genetic data.
2a: Humans begin with a genetic baseline: certain inclinations and reflexes, but with a lot to improve upon.
3a: Machines begin with the ability to give an output for an input, which at this point, are practically random.
Learning & Testing
1b: Life is tested through survival. Genes lead to physical features in addition to behaviors which can help or hinder.
2b: Humans experience the world. Early on as infants, we have a great inclination to explore and try out new things. Different actions lead to different consequences and we associate them together. Which materials are hard and shouldn’t be hit against. Which foods taste good or bad. Which things break when they’re dropped.
3b: Machines are trained with training data sets. Validated with validation data sets. Then tested with testing sets. A bunch of different algorithms (whatever they are, they are unknown to us) are run on it and graded on which has the best result.
Iteration
1c: Once a new generation of organisms come into existence, they are slightly different and hopefully better than the last. They then go through the same process as the previous ones did.
2c: Every repeated experience is an opportunity for a human to learn something more. Maybe if you lightly tap the wall it won’t hurt, maybe this food will taste better because it smells different, maybe an object won’t break if it is placed lightly on the ground
3c: The program with the best result is selected for a new batch that builds off of it.That batch is tested again and the process repeats.
Subtleties & Differences
Of course, the details in all of these aren’t exactly the same, and these parallels only serve as a very high-level connection among these processes. One key difference is the amount of time each process takes. From evolution which takes millions of years, to human learning which is continuous throughout a single lifetime, to machine learning which happens in moments. It seems the trend is a quickening of pace.
A fun note: The term “meme” was originally coined as a counterpart of the word gene. A knowledgeable equivalent to genetic data. It’s what makes humans different from other species. Rather than waiting for evolution to change, we accumulate and pass down knowledge which is much more malleable and subject to improvement.
Looking into the future, then comes the question of how machines can improve upon human performance. Are machines just good at their own thing? Or can they do everything a human can, but better? For example, computers are unequivocally better at most humans at doing quick arithmetic without error. However, can humans be better at reading emotions? Or writing music? Or are our strengths in different areas?
Another difference is what is considered good or bad. Though the goal might be the same, there are different ways to approach it. With evolution, the goal is survival which can lead to some interesting organisms that die immediately after reproducing or simply brute force reproduce as many offspring as possible. Humans can have vastly different priorities, but are all aiming to fulfill their basic needs and perhaps self-actualization afterwards. What are machines looking for? The “right” answer or most “accurate” prediction, or whatever we tell them, I suppose. Or rather, what they make of what we tell them, which may or not be exactly what we intended.
Nevertheless, all of these processes continue on. And with the changing circumstances, defining and redefining their goals, iterating and improving and redefining again.
Seeing these similarities, it’s fascinating and quite entertaining to ponder on how seemingly different processes are connected. What we make is a reflection of ourselves, and we are a reflection of the world.
