Deep Learning for Activity Recognition (A Brief and Incomplete Survey). Taylor. Workshop on Gesture Recognition (talk) 2011


http://videolectures.net/gesturerecognition2011_taylor_tutorial/

<Only 30 mins so probably pretty surface level>

  1. <Intro on common approach was not understandable to me>
  2. But the basic point is to use nets to develop feature detectors instead of hand-designing
  3. Tend to learn one level at a time in deep nets, unsupervised
  4. Often nets are used for classification but doesnt have to be
  5. Because representations are distributed they are more flexible than clustering techniques (k means is common in this field) which is based more on exemplars
  6. Convnets (biological motivation):
    1. Learn increasingly abstract representations as levels go up
    2. Feature maps <-> simple cells
    3. Pooled maps <->complex cells
  7. For video, can do 3D convnets (2D images + time)
    1. Pooling done across frames
  8. Lots of different architectures for doing deep learning for activity recognition.  Those discussed are very complex, and have differing levels of learned vs hard wired filters
    1. End layer is 128D
  9. Another approach is to try and construct a generative model of video
    1. Discusses “gated restricted Boltzmann machines” for this
    2. Extract temporal features, can be sensitive to motion, but also extracts static features. does a sort of segmentation
    3. Space-time deep belief networks
    4. Can be used to fill in missing pixels, or denoising
  10. “Stacked convolutional independent subspace analysis”
  11. Still of equivalent performance for more traditional methods in this field
  12. Room for progress is relying more on larger amounts of unsupervised data
  13. Work from Hinton questions doing invariant learning; instead treat them as latent features
Advertisements

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: