Why do the nodes in the final hidden layer of a neural network

>>59279367

>there is only one form of Network

nigga there are so many approaches and ways to do it its hard to keep track off .. educate yourself before you shitpost here

Depends on your network, some problems may require more than one output, ex: a network that learns a Mario level may want either one output saying what button to press during that frame, or maybe two outputs so that you can have two buttons pressed at a time.
It all depends on how you want the neural net setup

>>59279405
Yeah, but wouldn't it be other nodes determine which other outputs to press?

Think of the neural net as natural selection and the outputs as evolutionary branches. They all have undergone the same kind of heuristics, but with different random variables producing different results which have different benefits. It's possible that no single branch is "the best" and each has their own benefits.

It's also guaranteed that all of them are going to far shittier and have taken far more work to produce than if a skilled human actually sat down and worked on the problem.

>>59279367
your outputs would be bits in a lot of networks and 2 output is only 2 bits

>>59279444
So is it missing the point to try to understand what each node actually "does" in the AI? i.e., there is no node for one specific task, and it's all just a black box of wiring that doesn't make sense to humans?

>>59279434
>>59279510
Bump for answers to these. You know, maybe this belongs in a /sqt/ but I didn't think of it at the time and also it would probably just get buried.

>>59279510
Pretty much. Neural nets are made for tinkerers, not engineers.

>>59279510
>>59279832
>>59279367

Suppose we're trying to learn how to recognize numbers. A simple approach would be to have one node for each pixel on the left side, then 10 output nodes. When training the network, we use a 10-dimensional vector as our output. Suppose we have a picture of 0 for our training sample. Its label would be [1 0 0 0 0 0 0 0 0 0]. When we're done training, our goal is to get those 10 nodes as close to that vector as possible. We might have one output node for a task where there's a continuous spectrum we need to learn, but in this case there's 10 separate unrelated categories - it's better to use multiple outputs here.