Neural Networks( backpropagation) -
suppose have trained neural network....
my question same neural network generate data if apply output in present input?
i working on mnist dataset , wondered happen if train our network output side(using final output input side itself) using backpropagation algorithm.
it might trivial 1 still need community
my thinking says can data back( or approximations original dataset )... can justified?
as far know. can't. because activation functions (mostly) non-linear.
a neural network black box (see this answer). second of all, take f(x) = x^2. if want compute n f(n), there 2 possible solutions; same works neural networks, there can multiple solutions, it's impossible inverse of them. main point being: because know inverse of function, doesn't mean know inverse of neural-network. it's black-box!
however, can visualise response neuron gives input. example, 'aspects' neural network looks recognize face:
google deepdream amplifies aspects it's looking recognize objects. check out!

Comments
Post a Comment