A Committee of One - Using Dropout for Active Learning in Deep Networks
Master thesis
Permanent lenke
http://hdl.handle.net/11250/2352342Utgivelsesdato
2015Metadata
Vis full innførselSamlinger
Sammendrag
In many of the problem domains typically tackled by deep learning, datais plentiful and cheap but labeling of the data is tedious and expensive.Letting a model actively select the data instances it is uncertain about to trainon and ignore others can reduce the percentage of instances that mustbe labeled to achieve satisfactory results. To this end, this project presents anovel semi-supervised active learning algorithm called Active Deep Dropoutnetworks (ADD-networks). It is based on evaluating a deep neural network suncertainty on unlabeled instances, through measuring disagreement withina committee of networks derived from the original network. The committeemembers are Monte-Carlo-sampled from the full network using the conceptof dropout. Experiments on classifying handwritten digits show that ADD-networks are comparable to a state-of-the-art method, and vastly outperformsrandom selection of instances.