|
In one of his recent talks Andrew Ng mentioned several state-of-the-art results:
However, there were no references, and his online list of publications does not seem to include anything about learning on video. Any idea where I can read more about these? On a slightly separate subject, are people interested in collaborating on maintaining a comprehensive list of ML benchmarks and the most interesting results on them (like this slide, but more comprehensive and with more detail)? Should this be a table on Wikipedia? Edit: fixed slide, emphasis added |
|
I'm not sure if this is what you are asking about but I'd start with Ivan Laptev's work, which apparently was the previous state of the art on Hollywood2. Or are you asking specifically about the stanford feature learning results? To the second question: This would indeed be very helpful. I think there have been some efforts but I don't know of anything that really "made it". Though I feel this shouldn't be too hard if enough people collaborate. Yes, I'm interested in finding references for the four "video" results. Perhaps they weren't published by Andrew Ng, but by someone else at Stanford?
(Sep 12 '11 at 16:32)
Oleg Trott
|
|
I think "Stanford Feature Learning" is a branded name for a variety of techniques used in recent publications on unsupervised feature extraction by members of Andrew Ng's lab. I think a majority of them are based on deep learning models (e.g. Deep Belief Networks or Stacked Denoising Autoencoders with a sparsity constraint) with or without convolutional variants & pooling when it makes sense (e.g. for 1D audio and 2D images). Some of them are not even "deep": the CIFAR entry is probably referring to recent work by Adam Coates on using k-means centers or random samples carefully normalized / whitened and then used as a dictionary for a simple thresholed dot-product as sparse encoder. |
