Deep Learning Using Tiny Domain-Specific Datasets with Sparse Labels

Machine learning is an ever-expanding field of research, and recently deep learning has been the architecture of choice. However, traditional deep learning methodologies require substantial amounts of data to train their networks. This requirement for large data means that there are large numbers of real-world problems that cannot utilise the power of these deep learning networks due to a lack of data. Being able to use deep learning architectures with tiny domain-specific datasets would allow sectors such as healthcare to use deep learning as an aid in training and potentially in real time procedures. In this thesis, deep learning using tiny domain-specific datasets with sparse labels is achieved on two machine learning problems: semantic segmentation and action recognition. This is accomplished by utilising semi-unsupervised learning to train a convolutional neural network (CNN) to predict superpixels for an image, using a novel structural representation named Multi-channel Connected Graphs (MCGs). These deep-learned superpixels are then used in an end-to-end network consisting of two hourglass modules, with each specialising in a separate task; 1) deep learned superpixels, 2) semantic segmentation. For multi-class semantic segmentation, a variation on transfer learning is used. Action recognition with tiny amounts of training data is obtained by drastically reducing the input feature from full HD resolution down to $32\times32\times2$, with each cell consisting of the majority class in the first channel, and the secondary class in the second channel. This input is then used in a recurrent neural network consisting of multiple CNNs and bidirectional long short-term memory layers