Why neural stack is used?

Question

I followed the tutorial of neural stack in https://iamtrask.github.io/2016/02/25/deepminds-neural-stack-machine/. Its topic is "Learning to Transduce with Unbounded Memory", issued by Google DeepMind.

I roughly understood this proposed model (above website had easily explained!), but I have not read any other reference journals.

The last example of tutorial is making input sequence backward. However, I wondered why we just separate input sequence into words and reorder them backward using simple conditional loop? (And that would not be a neural network)

I mean why creating hidden layers and operating lots of equations just to make input sequence backward? Is there any advantage of using neural network?

I will read reference journals anyway. But for now, I just want to know the reason to use neural network, instead of programming with simple conditional loop; its advantages.

Answer 1

The point of neural networks / machine learning is to apply a general algorithm to a wide variety of problems, and using data and a valuation function, to effectively produce the desired output without you actually having to code it yourself.

In the case of reversing a stack; sure, it's easier to do this using normal code. But the point is that you would use a general algorithm that can do this without you actually having to code a stack explicitly. In a sense, machine learning solutions 'write themselves' based upon training data. The stack is a trivial example to show you how this works.