A geometric approach to active learning for convolutional neural networks