Hacker News Reader

Comment by notshift | Hacker News Reader

> for example, through an instant rewiring of the brain that doesn't require any practice.

You might have an alien species that can copy the brain state of someone who already knows a task, or one that can gain skills extremely quickly through practice, but I don't think practice is going away. The fact that neural networks have been the only way we've been able to solve a lot of problems, which approximate the way that the human brain learns, is pretty strong evidence of this. You start out with some weights, you measure loss, you adjust weights, and then you try again.

Justice as an emotion might go away or exist in a different form, but the underlying reason why humans have a sense of justice is evolutionary psychology / game theory. Probably any life form which is shaped by evolutionary forces would have some similar instinct. (Certainly not all possible intelligent life forms though, I'd agree.)

View on HN

> The fact that neural networks have been the only way we've been able to solve a lot of problems, which approximate the way that the human brain learns, is pretty strong evidence of this.

Chances are a lot of machine learning implementations can be replaced with just "boring" statistical models and achieve more power.

OTOH plenty of creatures don't need to learn. Does a mosquito need to learn? No, it spawns thousands of offspring and doesn't live very long. The high spawn rate means you have a wide variety of natural mutations in your offspring, meaning one or a few of them are likely to have higher fitness in a given niche. It doesn't matter if most die if a few go on to survive. This is the strategy many organisms use to dominate the world in far greater numbers than our own species.

I sure hope we don't have mass-reproducing space mosquitos in our future.

It's OK, space has too much radiation for things to migrate off world unshielded. Except if you are a tardigrade though, but the reasons for them being like that are due to the types of niches they occupy on earth. Mosquitos on earth that have some sort of space proof shielding in their exoskeleton would probably be quickly outcompeted by those more fit mosquitos that don't have to invest resources into this space proof exoskeleton.

> Chances are a lot of machine learning implementations can be replaced with just "boring" statistical models and achieve more power.

This is of course true, but the innovation in ML is not that a neural network (or whatever model) is equivalent to something else, it's finding the weights in the first place.

> The fact that neural networks have been the only way we've been able to solve a lot of problems, which approximate the way that the human brain learns, is pretty strong evidence of this

These are all difficult assertions to make because we're trying to prove a negative. Your claim is evidence that neural networks are one way to do it, and nobody's arging that. But it's not evidence that there doesn't exist a better way, or that alien life might evolve a different and even less optimal way.

Legs are pretty ubiquitous, even flying insects have them. If we'd never seen a fish or snake we might conclude that they're inevitable. Somebody that evolved to be rad-hard on a planet without a magnetosphere might conclude that life can't exist without the thick carapace that they're made of and only look for planets rich in silicon and calcium.

I agree with the legs example, but neural networks are basically a mathematical construct, and they solve problems that any other intelligent species in our universe would also have to face. I think if a better mathematical construct for solving those problems were possible, the process of natural selection probably would've found it by now.

We know that evolution has limitations in how it explores the state space; sometimes certain new developments depend upon past developments. But it seems to me that the development of a brain would hit relatively few of these barriers.

Reply on news.ycombinator.com