>Why is there no "general" deep learning algorithm?!
Because "deep learning" is almost always just neural networks with many layers and some tweaks to the unit types to alleviate vanishing-gradient issues, trained by stochastic gradient descent. In rather theoretical terms, "deep learning" involves using gradient descent, and gradient-increasing tweaks, to search for a specific continuous circuit in a given space of continuous circuits. It's only as general as the hypothesis class you search through, which here is, again, continuous "circuits" composed out of "neuron" units of specific types, numbers, and arrangements.
Now, in computability terms, given a large enough (continuous or discrete) circuit, possibly a recurrent one, you can represent any computable function. However, in learning terms, that doesn't make a useful computable function at all easy to find in a very-high dimensional space.
eli_gottlieb|9 years ago
Because "deep learning" is almost always just neural networks with many layers and some tweaks to the unit types to alleviate vanishing-gradient issues, trained by stochastic gradient descent. In rather theoretical terms, "deep learning" involves using gradient descent, and gradient-increasing tweaks, to search for a specific continuous circuit in a given space of continuous circuits. It's only as general as the hypothesis class you search through, which here is, again, continuous "circuits" composed out of "neuron" units of specific types, numbers, and arrangements.
Now, in computability terms, given a large enough (continuous or discrete) circuit, possibly a recurrent one, you can represent any computable function. However, in learning terms, that doesn't make a useful computable function at all easy to find in a very-high dimensional space.