do you know any application beside pattern recog. worthe in order to implement Hopfield neural network model?
+2
A:
Recurrent neural networks (of which hopfield nets are a special type) are used for several tasks in sequence learning:
- Sequence Prediction (Map a history of stock values to the expected value in the next timestep)
- Sequence classification (Map each complete audio snippet to a speaker)
- Sequence labelling (Map an audio snippet to the sentence spoken)
- Non-markovian reinforcement learning (e.g. tasks that require deep memory as the T-Maze benchmark)
I am not sure what you mean by "pattern recognition" exactly, since it basically is a whole field into which each task for which neural networks can be used fits.
bayer
2009-06-01 22:36:26
Could you be a bit more precise about their use with non-markovian reinforcement learning? You say they're used for tasks that require deep memory, but wouldn't multi layer backprops network do the same? What makes those best for the job? Thanks
devoured elysium
2009-11-23 14:15:25
Typically, you present the observations of the k last timesteps to an MLP to map it to an action or a value. That is something you can call "k-markovian". However, "deep memory" refers to situations when the number of timesteps you have to look back is not necessarily bounded. In that case, you need recurrent network which can (at least theoretically) look back in time an infinite number of timesteps. You might want to check this paper: http://www.idsia.ch/~daan/papers/jof.pdf
bayer
2009-11-24 10:19:36