Active learning method for high dimensional image data using transfer learning to reduce dimensionality

Deep learning algorithms that perform very well on a wide variety of tasks usually require a large amount of labelled samples. The labelling process especially for complex problems such as medical applications require expert knowledge making it costly and time consuming. To lower the amount of needed labelled samples the field of active learning aims to select the most valuable samples. Traditional active learning methods do not perform well for deep neural networks when there are still few labelled samples available because they rely on the output of the trained model which is unreliable. A method is proposed that maps the samples to a feature space using a pre-trained convolutional network, then clusters the samples and finally chooses the most representative sample from each cluster. The method does not require initially labelled samples. The comparison to random selection of samples shows that representative sampling results in a better performance when training a classification model. Further analysis show that the performance depends on the model chosen for feature selection but almost all choices out perform random selection. The use of the approach is especially high when few samples can be labelled and decreases for higher number of labels samples which would favour a combination of representative sampling for the initial samples and traditional active learning methods afterwards.