Suppose we have continuous stream of data which length we cannot predict and discretize. Is there a type of neural network that can hold this stream and makes output based on the information stored in this stream?


Any neural network trained using some variant of online learning (e.g. stochastic gradient descent) will be able to do this. If the stream contains independent samples, then a feedforward network would work. If the stream contains sequence data with time dependencies that you want to model, then a recurrent network would be the tool of choice (trained using a method like backpropagation through time).

  • $\begingroup$ I want to get an output based on the sequence. For example series of pixels and the output must be true or false but want to make it with random picture size and do not want to resize them. Is your answer suitable for this task? $\endgroup$
    – Blake
    Jun 12 '16 at 7:53
  • $\begingroup$ People have shown that it's possible to classify images (e.g. mnist dataset) using recurrent networks, even processing as a sequence of pixels presented one-at-a-time (of course this is harder than seeing the whole image at once). If the different size images have different spatial scales, this would be translated to different temporal scales, which would be an extra challenge but maybe possible. $\endgroup$
    – user20160
    Jun 12 '16 at 8:07
  • $\begingroup$ Buffering the pixels to form a full image, then feeding whole image to network might be easier. Number of input units is fixed so if you don't rescale you'll have to zero pad. If different size images have different spatial scales, this invariance must be learned. 'Spatial transformer' networks may be useful. btw recurrent nets can handle variable length sequences. $\endgroup$
    – user20160
    Jun 12 '16 at 8:12

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.