Parallels: Computation and EvolutionDecember 21, 2022
Assuming the entirety of human race is nothing but a network of brains with sensory inputs that function in an environment by taking actions and responding to rewards, it’s not hard to see how it compares to reinforcement learning.
At each point in time since our existence on this planet, the brain needed to solve various problems one at a time.
One of the first problems the brain encountered was survival. This not only included hunting for sustenance, making tools to optimise for ease of access to food, but also gave birth to the idea of settlements which eventually led to civilisation. All of this could be considered as the brain solving an optimisation problem of survival.
Just like how we would train an RL agent over an epoch of say, 10,000 steps, we as a network of brains have taken various paths in an effort to understand our environment better and each of this path led to various outcomes which not only includes wars, languages, but also gave birth to culture as we know it. Different paths to a similar outcome of survival and sustenance needed a network of brains to work together and help one another at times which led to various groups of them evolving over time which gave birth to culture.
Human brain is a complex biological machine that is capable of maximising the feeling of it’s sensory organs, some call it hedonism, but fundamentally it’s just an organism that’s trying to make sense of it’s environment within the bounds of it’s input organs — sense of touch, smell, sound, sight. Over time, we have explored each of this domain to various degrees and we continue to do so, It's not hard to see how some of these endeavours could fundamentally be perceived as art over time.
History as we know it, is quite simply a log of each of the training runs that has occurred since we’ve come to be. Just like how we learn from various training runs of an RL experiment, we do sometimes tend to learn from history — sometimes not.
The network of human brains (our species) is at a point where it has explored various routes to fulfil its objectives of getting to know its environment. Traditionally with RL, we have exploration and exploitation as two methods to learning the environment and solving for an optimisation function — looking at our own training run so far of human evolution, we quickly switched over to exploitation to understand our environment but we may have over done it to a point where we’re going back to dealing with the problem of survival. One could also say, we’ve come full circle starting with solving for survival and back to trying to solve for the very same thing.
Assuming these aren't overly coincidental or far fetched, it's super fascinating to think about these parallels that exist between these two very different domains of Biology and Computation.