|
I was just wondering about a scenario where you have two neural networks with 10 input neurons each (or "sensor neurons", whatever you'd like to refer to them as). What does it mean when they say that one of the nets takes continuous sensor input, and the other takes discrete? I understand what the difference between discrete and continuous are (in a numerical and mathematical sense), but I'm not sure how that relates to machine learning and neural networks, especially for what it means to say that a network "takes in" continuous (or discrete) data. Any pointers or insights would be useful and appreciated! Thank you |
|
For multilayer perceptrons, as far as I can tell, the only difference is in how you encode the features you pass to the neural network. Real-valued features are usually encoded as real numbers (but they don't have to be, you can bin them, etc), while discrete features are usually encoded with one-hot binary units (that is, a subset of the input neurons, only one of it being assigned the value 1 while the others are assigned the value 0 or -1). As far as the optimization algorithms are concerned nothing changes. For other kinds of neural networks things are different. Gaussian restricted boltzman machines, for example, behave quite differently from logistic RBMs. |
