If neural networks appear to be the most promising approach to creating more sophisticated AI like we have seen...

yeah thats true but it has a bit more depth than some standard anns. and i think if you read some of eliasmiths book it at least gives some info about why you should buy that we can use spiking neural networks (ie differential equations with random nonlinearities btwn) with some sort of plausibility and some stuff on why you can structure things certain ways. the model itself i dont care that much about, i just think its moving somewhat in the right direction in certain parts

>(ie differential equations with random nonlinearities btwn)
You clearly don't know what you're talking about.

true maybe idk man

It something beyond that, not only ANNs. We're really starting to shift into some sort of "implicit knowledge", where problems arent solved analitically anymore so we can reach out better solutions thanks to computers

Do you think a BFS approach to research is ineffective in general, or are you just not convinced of this line of thinking?

Also, what about the approach by numenta.com/?

Evolution has been going on for many hundred million years, to not take advantage of already existing designs would be starting from scratch.

Take the submarine, for example, or fighter formations that were based on bird formations that save 15% of fuel, or camouflage that animals have, these are methods that helped the animals to survive.

Why would not mimicking a brain not be a good idea?

What about Neuroevolution and algos like NEAT or HyperNEAT?

Neural networks are limited by the computational capabilities of computers. We can't accurately model a brain because our computers are too shit. So we have to make sacrifices and scale things down/simplify them for our shitty silicone chips.

en.m.wikipedia.org/wiki/OpenWorm?wprov=sfla1
We've managed to simulate a worm brain so maybe we can gain significant insights from that for neural networks

I still don't understand what a Neural Network is.

Hawkins seems a bit more motivated by the actual structure of the nervous system but his group really has yet to produce anything substantial beyond his anomaly detection engine. I used to follow his work pretty closely but it seems like things have petered out. I like his focus on on-line learning and continuous time-series prediction but I think you can make more progress by taking a more mathematically driven approach as opposed to a biological one. Ultimately though I think his goal is more to understand the brain and derive a system based on it than to create something that works but is not biologically realistic, so judging progress must be more nuanced than simply measuring accuracy of this or that benchmark.

Look up 3blue1brown on the youtubes

>We've managed to simulate a worm brain
No we haven't. Just take a look at the group's admitted progress:
github.com/openworm/OpenWorm/milestones

>en.m.
>m
Probably explains your lack of critical thought.

In their most basic form as a multilayer perceptron:
[math]W_n \cdot f_{n-1}(W_{n-1} \cdot \ldots \cdot f_1(W_1 \cdot x + b_1) + \ldots b_{n-1}) + b_n[/math]

Things have gotten more complicated but this is still the basic idea.

Well at least we are trying so there are scientists who believe simulating a worm brain is worthwhile.
Also I hope you don't have a phone because that would clearly mean you're a brainlet

>github.com/openworm/OpenWorm/milestones
>5 open 25 closed

83% seems pretty good

So, lets take a step back and think about what neural networks are actually doing.

They allow you to learn differentiable programs from data. In the case of a feedforward network, you can learn a fixed length program (the number of layers in the network). In a recurrent network you can learn arbitrary length programs.

So we have neural networks which can program themselves as long as we give them the proper input/output pairs. This seems like it will be a very useful part of an AI system, but its clearly not going to bring us all the way.

To me, the next step is to start thinking in terms of Model/Actor/Critic systems. We basically know how to do Actor/Critic, just look at the reinforcement learning literature, and we can do this with neural nets. The hard part is the model. How do we learn a model of the world with sparse, or non-existent reward signals? This is the current grand challenge of AI.

Look at a system like AlphaGo. The actor/critic aspects are fairly simple. The "Model" in alpha go is basically the Tree Search component. In a game like go, you don't have to learn a world model, you can get away with a basic tree search setup, and it is competely sufficient. When a human tries to play a game like Go or Chess, we have to learn to do a kind of tree search after playing the game for some time. Playing a head a few moves in your head is your model of the game. We need AI systems that can learn to build these types of models automatically, for arbitrary "games" or environments.

>I don't know how software development works.

This is how mimicking evolution could help, let populations of neural networks breed,"mutate",etc and change their structure

>biological vs mathematical
I cannot form an argument in full agreement or disagreement. I agree that, for the foreseeable future, mathematical approaches will produce more fruit. However, I think our attempts to mimic nature will increase over time.

Using nature as a model, however roughly, has been a thread more or less since the beginning. Perhaps the thread is strengthening? Perhaps the trend to try to understand the networks in the literature is evidence? Hinton's capsules are further evidence? I don't know.