ConvNet Features off-the-shelf: an Astounding Baseline for Object Recognition

Quick summary for people who don't follow machine learning and computer vision:

A team from NYU recently open sourced their neural net model from a recent object recognition competition. The team from KTH used that pre-trained net as a feature extractor and applied it to other standard datasets by learning an SVM on the extracted features. This "simple" approach did as well if not better than most state of the art methods, which were hand engineered for the specific tasks.

CNNs are great, but so prone to overfitting, thankfully that last few years have been good to the deep learning community - dropout, jitter, etc. - in combating these problems. Hopefully in the next few years more books are published to help dumb-down the material since reading a lot of these papers (in terms of trying to implement/expand/critique the material - just the history alone is interesting - hopfield nets, perceptrons, hebbian learning, ... ) is still very tough as results are difficult to reproduce

There is a nice python wrapper, called nolearn [1] that uses a pre-trained CNN called DeCaF to extract features prior to the final classification layer. I found the results to be surprisingly good in a lot of situations.

[1]: https://pythonhosted.org/nolearn/

I have to wonder - are there some techniques to build the trained models by crowdsourcing?