On Monday, 27 June 2016 at 22:17:55 UTC, Martin Nowak wrote:
On 06/27/2016 08:13 PM, Guillaume Piolat wrote:
Not yet, but it could be useful for new types of audio effects
and specific tasks like voiced/unvoiced detection.
There are many simpler solutions for that than using machine
learning. Writing a simple neural network with backpropagation
is fairly trivial, if you had that in mind to emulate existing
audio effects, not sure if it works well though.
I could probably write a simple backpropogation one, but I would
probably screw something up if I wrote my own convolutional
neural network.
My suggestion is that anyone interested in deep learning might
want to break it up into some more manageable projects.
For instance, one of the features of TensorFlow is
auto-differentiation. This means that you can provide it
arbitrary functions and it will calculate the gradients for you
exactly instead of relying on numerical estimates. autodiff
involves building an AST for a function and then walking it to
generate the gradient. A D autodiff library would probably be
easier to write than a comparable one in other languages since it
could take advantage of all the compile time functionality.
Alternately, TensorFlow also works well with heterogeneous
systems, so any work that improves D's capabilities with
OpenCL/CUDA or MPI would be something that might make it easier
to develop a D Deep Learning library.
