d-SNE: Domain Adaptation Using Stochastic Neighborhood Embedding

Abstract

Deep neural networks often require copious amount of labeled-data to train their scads of parameters. Training larger and deeper networks is hard without appropriate regularization, particularly while using a small dataset. Laterally, collecting well-annotated data is expensive, timeconsuming and often infeasible. A popular way to regularize these networks is to simply train the network with more data from an alternate representative dataset. This can lead to adverse effects if the statistics of the representative dataset are dissimilar to our target. This predicament is due to the problem of domain shift. Data from a shifted domain might not produce bespoke features when a feature extractor from the representative domain is used. In this paper, we propose a new technique (d-SNE) of domain adaptation that cleverly uses stochastic neighborhood embedding techniques and a novel modified-Hausdorff distance. The proposed technique is learnable end-to-end and is therefore, ideally suited to train neural networks. Extensive experiments demonstrate that d-SNE outperforms the current states-of-the-art and is robust to the variances in different datasets, even in the one-shot and semi-supervised learning settings. d-SNE also demonstrates the ability to generalize to multiple domains concurrently.

Few-Shot Learning for Ct Scan Based Covid-19 Diagnosis

“…In this sub-section, we focus on comparing the performance between the proposed method and other state-of-the-art supervised domain adaptation methods, including CCSA [11], FADA [10], d-SNE [12]. In order to show how does our domain adaptation help to improve the cross-domain diagnostic performance, we also involve a source only competitor which is trained on a deep network with only source domain data and tested on a target domain test set.…”

“…CCSA [11] proposed a series of loss functions in order to manage the domain gap for a few-shot domain adaptation tasks. d-SNE [12] introduced a new approach that exploits the stochastic neighborhood embedding theory and modified-Hausdorff distance to improve the few-shot classification performance. Although, many efforts have been done on SDA or few-shot COVID-19 diagnosis areas [13,14], applying domain adaptation on CT images for the COVID-19 diagnostic task is relatively a new area, and our proposed method is one of the first attempts in utilizing synthetic chest CT scans for fewshot COVID-19 diagnostic task.…”

Unsupervised Domain Adaptation in Activity Recognition: A GAN-Based Approach

“…As labels are available in supervised domain adaption, it is possible to perform within-class adaptation. For example, Xu et al [45] propose d-SNE where samples from both source and target domains are transformed to common latent space; i.e., stochastic neighborhood embedding (SNE) space, and then a modified Hausdoff distance is employed to minimise the distance between samples from the same classes but maximise the distance between samples from different classes. Morsing et al [26] propose to deal with covariate shift by connecting samples in a penalty graph structure.…”

Adversarial Representation Learning for Text-to-Image Matching

“…Text-Image Matching: Learning cross-modal embeddings has numerous applications [61,69] ranging from PINs using facial and voice information [37], to generative feature learning [15] and domain adaptation [63,65]. Nagrani et al [37] demonstrated that a joint representation can be learned from facial and voice information and introduced a curriculum learning strategy [3,45,46] to perform hard negative mining during training.…”