|
Can anyone point me to recent unsupervised algorithms which are similar to Slow Feature Analysis (SFA) and ICA where high-level features are extracted from spatially correlated time series data (similar to movies but not exactly). |
|
I think, Deep Learning from Temporal Coherence in Video might be an interesting start. Also, all future papers that reference it (still managable :)), e.g. Slow, decorrelated features for pretraining complex cell-like networks from Bergstra et al. The temporal coherence constraint looks very interesting. Right now, I've been working in frequency space using PSD. I suspect the poor results I'm getting working directly with time is phase problems. The data is recorded at an unknown state. The method they propose seems to treat each point in the time series as separate examples.
(Nov 15 '11 at 00:34)
crdrn
|
|
There are some connectionist models like this: "Unsupervised feature learning for audio classification using convolutional deep belief networks" There is also some work from the Hinton group working with time series mocap data, I don't know the title of the paper though. There is also work on space-time features and interest points at Ivan Laptev's lab. I would not really call them "high level" but they seems to work very well for movie data.
This answer is marked "community wiki".
|
The exact type of time series data matters a LOT, if we're to make helpful reccommendations. What is it?
fMRI captures with time between each volume being ~400ms (for those familiar, resting state captures TR ~= 400ms).