Generalization to out-of-distribution (OOD) data is a capability natural to humans yet challenging for machines to reproduce. This is because most statistical learning algorithms strongly rely on the i.i.d.~assumption while in practice the target data often come from a different distribution than the source data, known as domain shift. Domain generalization (DG) aims to achieve OOD generalization by only using source domain data for model learning. Since first introduced in 2011, research in DG has undergone a decade progress. Ten years of research in this topic have led to a broad spectrum of methodologies, e.g., based on domain alignment, meta-learning, data augmentation, or ensemble learning, just to name a few; and have covered various applications such as object recognition, segmentation, action recognition, and person re-identification. In this paper, for the first time, a comprehensive literature review is provided to summarize the ten-year development in DG. First, we cover the background by giving the problem definitions and discussing how DG is related to other fields like domain adaptation and transfer learning. Second, we conduct a thorough review into existing methods and present a taxonomy based on their methodologies and motivations. Finally, we conclude this survey with potential research directions.

Understanding relationships between feature variables is one important way humans use to make decisions. However, state-of-the-art deep learning studies either focus on task-agnostic statistical dependency learning or do not model explicit feature dependencies during prediction. We propose a deep neural network framework, dGAP, to learn neural dependency Graph and optimize structure-Aware target Prediction simultaneously. dGAP trains towards a structure self-supervision loss and a target prediction loss jointly. Our method leads to an interpretable model that can disentangle sparse feature relationships, informing the user how relevant dependencies impact the target task. We empirically evaluate dGAP on multiple simulated and real datasets. dGAP is not only more accurate, but can also recover correct dependency structure.

Speech separation has been shown effective for multi-talker speech recognition. Under the ad hoc microphone array setup where the array consists of spatially distributed asynchronous microphones, additional challenges must be overcome as the geometry and number of microphones are unknown beforehand. Prior studies show, with a spatial-temporalinterleaving structure, neural networks can efficiently utilize the multi-channel signals of the ad hoc array. In this paper, we further extend this approach to continuous speech separation. Several techniques are introduced to enable speech separation for real continuous recordings. First, we apply a transformer-based network for spatio-temporal modeling of the ad hoc array signals. In addition, two methods are proposed to mitigate a speech duplication problem during single talker segments, which seems more severe in the ad hoc array scenarios. One method is device distortion simulation for reducing the acoustic mismatch between simulated training data and real recordings. The other is speaker counting to detect the single speaker segments and merge the output signal channels. Experimental results for AdHoc-LibiCSS, a new dataset consisting of continuous recordings of concatenated LibriSpeech utterances obtained by multiple different devices, show the proposed separation method can significantly improve the ASR accuracy for overlapped speech with little performance degradation for single talker segments.

The literature on robustness towards common corruptions shows no consensus on whether adversarial training can improve the performance in this setting. First, we show that, when used with an appropriately selected perturbation radius, $\ell_p$ adversarial training can serve as a strong baseline against common corruptions. Then we explain why adversarial training performs better than data augmentation with simple Gaussian noise which has been observed to be a meaningful baseline on common corruptions. Related to this, we identify the $\sigma$-overfitting phenomenon when Gaussian augmentation overfits to a particular standard deviation used for training which has a significant detrimental effect on common corruption accuracy. We discuss how to alleviate this problem and then how to further enhance $\ell_p$ adversarial training by introducing an efficient relaxation of adversarial training with learned perceptual image patch similarity as the distance metric. Through experiments on CIFAR-10 and ImageNet-100, we show that our approach does not only improve the $\ell_p$ adversarial training baseline but also has cumulative gains with data augmentation methods such as AugMix, ANT, and SIN leading to state-of-the-art performance on common corruptions. The code of our experiments is publicly available at https://github.com/tml-epfl/adv-training-corruptions.

Code-switching is a common phenomenon among people with diverse lingual background and is widely used on the internet for communication purposes. In this paper, we present a Recurrent Neural Network combined with the Attention Model for Language Identification in Code-Switched Data in English and low resource Roman Urdu. The attention model enables the architecture to learn the important features of the languages hence classifying the code switched data. We demonstrated our approach by comparing the results with state of the art models i.e. Hidden Markov Models, Conditional Random Field and Bidirectional LSTM. The models evaluation, using confusion matrix metrics, showed that the attention mechanism provides improved the precision and accuracy as compared to the other models.

The key challenge of unsupervised vehicle re-identification (Re-ID) is learning discriminative features from unlabelled vehicle images. Numerous methods using domain adaptation have achieved outstanding performance, but those methods still need a labelled dataset as a source domain. This paper addresses an unsupervised vehicle Re-ID method, which no need any types of a labelled dataset, through a Self-supervised Metric Learning (SSML) based on a feature dictionary. Our method initially extracts features from vehicle images and stores them in a dictionary. Thereafter, based on the dictionary, the proposed method conducts dictionary-based positive label mining (DPLM) to search for positive labels. Pair-wise similarity, relative-rank consistency, and adjacent feature distribution similarity are jointly considered to find images that may belong to the same vehicle of a given probe image. The results of DPLM are applied to dictionary-based triplet loss (DTL) to improve the discriminativeness of learnt features and to refine the quality of the results of DPLM progressively. The iterative process with DPLM and DTL boosts the performance of unsupervised vehicle Re-ID. Experimental results demonstrate the effectiveness of the proposed method by producing promising vehicle Re-ID performance without a pre-labelled dataset. The source code for this paper is publicly available on `https://github.com/andreYoo/VeRI_SSML_FD.git'.

With the rapid development in sensor and measurement technology, time-series data of processes have become available in large amounts with high accuracy. Machine learning and data science play an important role in analyzing and perceiving information of the underlying process dynamics from these data. Building a model describing the dynamics is vital in designing and optimizing various processes, as well as predicting their long-term transient behavior. Inferring a dynamic process model from data, often called system identification, has a rich history; see, e.g., [30,46]. While linear system identification is well established, nonlinear system identification is still far from being as good understood as for linear systems, despite having a similarly long research history, see, e.g., [25, 44]. Nonlinear system identification often relies on a good hypothesis of the model; thus, it is not entirely a black-box technology. Fortunately, there are several scenarios where one can hypothesize a model structure based on a good understanding of the underlying dynamic behavior using expert knowledge or experience. Towards nonlinear system identification, a promising approach based on a symbolic regression was proposed [4] to determine the potential structure of a nonlinear system.

Recently, practical brain-computer interface is actively carried out, especially, in an ambulatory environment. However, the electroencephalography (EEG) signals are distorted by movement artifacts and electromyography signals when users are moving, which make hard to recognize human intention. In addition, as hardware issues are also challenging, ear-EEG has been developed for practical brain-computer interface and has been widely used. In this paper, we proposed ensemble-based convolutional neural networks in ambulatory environment and analyzed the visual event-related potential responses in scalp- and ear-EEG in terms of statistical analysis and brain-computer interface performance. The brain-computer interface performance deteriorated as 3-14% when walking fast at 1.6 m/s. The proposed methods showed 0.728 in average of the area under the curve. The proposed method shows robust to the ambulatory environment and imbalanced data as well.

The auditory attention decoding (AAD) approach was proposed to determine the identity of the attended talker in a multi-talker scenario by analyzing electroencephalography (EEG) data. Although the linear model-based method has been widely used in AAD, the linear assumption was considered oversimplified and the decoding accuracy remained lower for shorter decoding windows. Recently, nonlinear models based on deep neural networks (DNN) have been proposed to solve this problem. However, these models did not fully utilize both the spatial and temporal features of EEG, and the interpretability of DNN models was rarely investigated. In this paper, we proposed novel convolutional recurrent neural network (CRNN) based regression model and classification model, and compared them with both the linear model and the state-of-the-art DNN models. Results showed that, our proposed CRNN-based classification model outperformed others for shorter decoding windows (around 90% for 2 s and 5 s). Although worse than classification models, the decoding accuracy of the proposed CRNN-based regression model was about 5% greater than other regression models. The interpretability of DNN models was also investigated by visualizing layers' weight.

Deep learning is the current darling of AI. Used by behemoths such as Microsoft, Google and Amazon, it leverages artificial neural networks that "learn" through exposure to immense amounts of data. By immense we mean internet-scale amounts -- or billions of documents at a minimum. If your project draws upon publicly available data, deep learning can be a valuable tool. The same is true if budget isn't an issue.