DGIST Team Introduces, ‘DRANet’, An AI Neural Network Module That Can Separate And Convert Environmental Information In The Form Of Complex Images Using Deep Learning

As a result of recent advances in Deep Learning (DL), deep learning neural networks (DNN) have been widely used to improve model performance in computer vision, natural language processing, and more. However, existing domain adaptation methods learn only associated features that share a domain. Thus domain gaps between data significantly degrade the existing model performance.

Unsupervised domain adaptation has been used in many studies to generalise models across domains. This method aligns the distribution shift in labelled source data to the distribution shift in unlabeled target data. Various strategies have been investigated to bridge the gap between domains, including feature learning and generative pixel-level adaptation.

Disentangling representations into exclusive and shared components in a latent feature space has advanced the field of study. Disentangling representations enhance a model’s capacity to extract invariant domain characteristics and its domain adaption performance. These approaches, however, continue to focus on associated features between two domains, such as shared and exclusive components. This compels the use of several encoders and generators specialised in different domains. In addition to domain classifiers, the network training strongly relies on a task classifier containing ground-truth class labels.

Researchers from the Department of Information & Communication Engineering, DGIST and POSTECH introduce a network architecture that disentangles image representations and transfers the visual attributes in a latent space for unsupervised cross-domain adaptation. DRANet is a single feed-forward network that does not require any ground-truth task labels for cross-domain adaptation. DRANet is the first approach based entirely on the individual domain characteristics for unsupervised cross-domain adaptation. 

Unlike existing methods, the proposed approach focuses on extracting the domain-specific features that preserve individual domain characteristics. Then it disentangles the discriminative features of individual domains into the content and style components using a separator, which is later used to generate the domain-adaptive features. 

Source: https://arxiv.org/pdf/2103.13447.pdf

The team explain that various domains may have different content and style distributions, and therefore just linearly separating the latent vector will not be enough. This inspired them to create a network design that uses non-linear separation and domain-specific scale parameters to handle such inter-domain differences. This architecture offers a multidirectional domain transfer from totally unlabeled data using a single encoder-decoder network.

Source: https://arxiv.org/pdf/2103.13447.pdf

In addition, they also propose a content-adaptive domain transfer module that aids in the preservation of scene structure while transferring style. When tested on standard digit classification and semantic segmentation tasks, the proposed network demonstrates SOTA performance. The results show that the network is capable of separating content-style factors and synthesising visually appealing domain-transferred images.

The team hopes that the proposed network will contribute significantly to future advancements in AI, such as image conversion and domain adaptation.

Paper: https://arxiv.org/pdf/2103.13447.pdf