|
Hello, What's the format of benchmark datasets used for RBM/DBN? Are they usually binary or real valued? For example, the input of Netflix prize data can be described as integer ranging from 1-5, along with date info, movie/user id. I'm curious about the following and any other benchmark i didn't mention, also any other image datasets. Do people use real-valued RBM? or convert the image to use binary RBM? MNIST: http://yann.lecun.com/exdb/mnist/ mentions that the dataset is binary: "The MNIST database was constructed from NIST's Special Database 3 and Special Database 1 which contain binary images of handwritten digits.", and I suppose images like this makes sense, as in each pixel is either black or white, so it is binary. But I've also seen images like this where there are background pixels where not all pixels are clearly not binary. In this case, do people use real-valued RBM or convert the dataset to be binary somehow? TIMIT: I'm aware that this is audio data, but since you have to pay to download it, I haven't had opportunity to look at it. Is the data format binary input or continuous number? Same question for NLP datasets. Thank you. |
|
MNIST is approximately binary, so usually binary RBMs are used. TIMIT: I am not familiar, but I suppose it has to be Gaussian visible units. NLP: not familiar, is this text? Then binary RBM. |
