Can pre-trained convolutional neural networks be directly used as a feature extractor for video-based neonatal sleep and wake classification?