Being able to tell the difference between large and small or faster and slower is an important feature evolved cognitive systems should have. We heard about value judgement in psychology and I can’t help myself thinking about this meme:
Sorry, I didn’t find a good “Monkey wants juice” reference…
Can we evolve a Markov Brain that possesses similar capabilities? I emphasize Markov Brain (MB) here because we know that ANNs already implement a ton of math without the need to evolve that in the first place. While the “can we evolve this” question is not necessarily a very scientific one, the constraints of that experiment might very well be, but let me explain the fitness landscape and environment first. We take a MB and either give it a retina that has 50/50 black and white dots (0 and 1) and we slowly flip these inputs randomly however we have a bias either towards ending up with 70/30 or 30/70 white/black. The MB should at some point indicate if it sees more black or more white dots. The same can be done with a single input that starts with a random frequency of black and white and slowly shifts this frequency to the target frequency.
The interesting question now are:
- Will brains that evolve to solve the retina task be very different from the single input one
- How does external constraints like rewarding accuracy over speed or the other way round effect performance, topology of the brain, evolutionary trajectory…
- How noise tolerant will the MB be
- Will the evolved topology even be remotely comparable to what we find in nature (for what we actually know about brains)
There are more complicated questions possible, but I will use them for the next project idea.