Unsupervised Learning of Useful and Interpretable Representations from Image Data

Sammanfattning: This master thesis tackles the problem of unsupervised learning of useful and interpretable representations from image data using deep Convolutional Neural Networks (CNN). Recent years have seen remarkable success from using deep learning technologies to tackle computer vision problems. This success is in part attributable to the availability of large, manually-annotated datasets; however, most image data is unlabelled and unstructured. It would therefore be beneficial to reduce the dependency on labelled datasets by training networks in an unsupervised manner. Ideally, we would like the extracted representations from such networks to be useful for a range of downstream machine learning tasks. Furthermore, we would like the learned representations to be interpretable in that the individual dimensions of the representation disentangle the true factors of variation in our image dataset. We trained state-of-the-art disentangled representation learning algorithms on synthetic and realworld datasets. We found that we could learn disentangled representations consisting of features with a one-to-one correspondence with known, groundtruth factors of variation in a synthetic dataset. Identification of disentangled models requires supervision in the form of labels or human evaluation. On a real-world dataset with unknown factors of variation we were able to learn highly informative representations using various unsupervised learning algorithms, some of which were especially robust to hyperparameter settings and random seeds. Identifying disentangled models required extensive human effort, and it was difficult to interpret the learned representations. Our results therefore suggest two main directions for future work: developing methods to identify disentangled models using minimal supervision, and developing novel methods for learning interpretable representations from real-world datasets.