Deep learning has proven to be beneficial for complex tasks
such as classifying images. However, this approach has been mostly applied
to static datasets.
Let's say we have a sequence of actions dataset. There are 10 different actions, but let say for simplicity that we have only a1 and a2 actions. The data are not stationary. For some time we have one distribution of actions sequences probabilities and then another. For my task RBM can model a distribution of stationary subsequences very well. However, I want to switch to new RBM as soon as distribution properties are changed. Also, I want to be able, after a training, to find a cluster(RBM?) to what this sequence of actions applies best. Is it realistic?
I have the raw idea that when the current RBM performance begins to degrade significantly we can switch to a new model. And we can chose the right RBM by the best performance if no one is not good enough(isn't clear how to measure this correctly) then - create new.
Do you have any ideas or hints, maybe some working alternatives?
If you would consider algebraic (not statistically based) segmentation of a non-stationary time series - you could take a look at R.Palivonaite, K.Lukoseviciute, M.Ragulskis. Algebraic segmentation of short nonstationary time series based on evolutionary prediction algorithms. Neurocomputing. 2013, vol. 121, p.354-364.
Hi Pavel,
If I understand your problem correctly, it is similar to the "weak sequence learning" which is sub topic of sequence learning problems that i am currently working on it.
as you want to modeling your problem with RBM models, you have two options for that:
1. using special variants of RBM models which is develop for time series data, such as Factored Conditional RBM
2. or if you have adequate sets of stationary time series sequence for each action, as you said, RBM can model a distribution of stationary sub-sequences, so you have a trained RBM model for each action, then it is possible enabling by soft-max function, to apply your second task ("what this sequence of actions applies best").
Note: option 2 works with Gated-RBM (they use extra units (gates) for switching between distributions)
I refereed you to GW Taylor works
Conference Paper Factored conditional restricted Boltzmann Machines for model...
Conference Paper Gated Softmax Classification.
https://www.researchgate.net/publication/221346352_Factored_conditional_restricted_Boltzmann_Machines_for_modeling_motion_style?ev=post
Conference Paper Factored conditional restricted Boltzmann Machines for model...
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Algorithms