In this paper, we propose a novel approach to solve the pose guided person image generation task. We assume that the relation between pose and appearance information can be described by a simple matrix operation in hidden space. Based on this assumption, our method estimates a pose-invariant feature matrix for each identity, and uses it to predict the target appearance conditioned on the target pose. The estimation process is formulated as a p-norm regression problem in hidden space. By utilizing the differentiation of the solution of this regression problem, the parameters of the whole framework can be trained in an end-to-end manner. While most previous works are only applicable to the supervised training and single-shot generation scenario, our method can be easily adapted to unsupervised training and multi-shot generation. Extensive experiments on the challenging Market-1501 dataset show that our method yields competitive performance in all the aforementioned variant scenarios.
In a news recommender system, a reader's preferences change over time. Some preferences drift quite abruptly (short-term preferences), while others change over a longer period of time (long-term preferences). Although the existing news recommender systems consider the reader's full history, they often ignore the dynamics in the reader's behavior. Thus, they cannot meet the demand of the news readers for their time-varying preferences. In addition, the state-of-the-art news recommendation models are often focused on providing accurate predictions, which can work well in traditional recommendation scenarios. However, in a news recommender system, diversity is essential, not only to keep news readers engaged, but also to play a key role in a democratic society. In this PhD dissertation, our goal is to build a news recommender system to address these two challenges. Our system should be able to: (i) accommodate the dynamics in reader behavior; and (ii) consider both accuracy and diversity in the design of the recommendation model. Our news recommender system can also work for unprofiled, anonymous and short-term readers, by leveraging the rich side information of the news items and by including the implicit feedback in our model. We evaluate our model with multiple evaluation measures (both accuracy and diversity-oriented metrics) to demonstrate the effectiveness of our methods.
Millions of people around the world are sharing COVID-19 related information on social media platforms. Since not all the information shared on the social media is useful, a machine learning system to identify informative posts can help users in finding relevant information. In this paper, we present a BERT classifier system for W-NUT2020 Shared Task 2: Identification of Informative COVID-19 English Tweets. Further, we show that BERT exploits some easy signals to identify informative tweets, and adding simple patterns to uninformative tweets drastically degrades BERT performance. In particular, simply adding 10 deaths to tweets in dev set, reduces BERT F1- score from 92.63 to 7.28. We also propose a simple data augmentation technique that helps in improving the robustness and generalization ability of the BERT classifier.
The ability to recognize human partners is an important social skill to build personalized and long-term human-robot interactions, especially in scenarios like education, care-giving, and rehabilitation. Faces and voices constitute two important sources of information to enable artificial systems to reliably recognize individuals. Deep learning networks have achieved state-of-the-art results and demonstrated to be suitable tools to address such a task. However, when those networks are applied to different and unprecedented scenarios not included in the training set, they can suffer a drop in performance. For example, with robotic platforms in ever-changing and realistic environments, where always new sensory evidence is acquired, the performance of those models degrades. One solution is to make robots learn from their first-hand sensory data with self-supervision. This allows coping with the inherent variability of the data gathered in realistic and interactive contexts. To this aim, we propose a cognitive architecture integrating low-level perceptual processes with a spatial working memory mechanism. The architecture autonomously organizes the robot's sensory experience into a structured dataset suitable for human recognition. Our results demonstrate the effectiveness of our architecture and show that it is a promising solution in the quest of making robots more autonomous in their learning process.
In the context of compressed sensing (CS), this paper considers the problem of reconstructing sparse signals with the aid of other given correlated sources as multiple side information. To address this problem, we theoretically study a generic \textcolor{black}{weighted $n$-$\ell_{1}$ minimization} framework and propose a reconstruction algorithm that leverages multiple side information signals (RAMSI). The proposed RAMSI algorithm computes adaptively optimal weights among the side information signals at every reconstruction iteration. In addition, we establish theoretical bounds on the number of measurements that are required to successfully reconstruct the sparse source by using \textcolor{black}{weighted $n$-$\ell_{1}$ minimization}. The analysis of the established bounds reveal that \textcolor{black}{weighted $n$-$\ell_{1}$ minimization} can achieve sharper bounds and significant performance improvements compared to classical CS. We evaluate experimentally the proposed RAMSI algorithm and the established bounds using synthetic sparse signals as well as correlated feature histograms, extracted from a multiview image database for object recognition. The obtained results show clearly that the proposed algorithm outperforms state-of-the-art algorithms---\textcolor{black}{including classical CS, $\ell_1\text{-}\ell_1$ minimization, Modified-CS, regularized Modified-CS, and weighted $\ell_1$ minimization}---in terms of both the theoretical bounds and the practical performance.
Though learning has become a core technology of modern information processing, there is now ample evidence that it can lead to biased, unsafe, and prejudiced solutions. The need to impose requirements on learning is therefore paramount, especially as it reaches critical applications in social, industrial, and medical domains. However, the non-convexity of most modern learning problems is only exacerbated by the introduction of constraints. Whereas good unconstrained solutions can often be learned using empirical risk minimization (ERM), even obtaining a model that satisfies statistical constraints can be challenging, all the more so a good one. In this paper, we overcome this issue by learning in the empirical dual domain, where constrained statistical learning problems become unconstrained, finite dimensional, and deterministic. We analyze the generalization properties of this approach by bounding the empirical duality gap, i.e., the difference between our approximate, tractable solution and the solution of the original (non-convex)~statistical problem, and provide a practical constrained learning algorithm. These results establish a constrained counterpart of classical learning theory and enable the explicit use of constraints in learning. We illustrate this algorithm and theory in rate-constrained learning applications.
With the ubiquitous graph-structured data in various applications, models that can learn compact but expressive vector representations of nodes have become highly desirable. Recently, bearing the message passing paradigm, graph neural networks (GNNs) have greatly advanced the performance of node representation learning on graphs. However, a majority class of GNNs are only designed for homogeneous graphs, leading to inferior adaptivity to the more informative heterogeneous graphs with various types of nodes and edges. Also, despite the necessity of inductively producing representations for completely new nodes (e.g., in streaming scenarios), few heterogeneous GNNs can bypass the transductive learning scheme where all nodes must be known during training. Furthermore, the training efficiency of most heterogeneous GNNs has been hindered by their sophisticated designs for extracting the semantics associated with each meta path or relation. In this paper, we propose WIde and DEep message passing Network (WIDEN) to cope with the aforementioned problems about heterogeneity, inductiveness, and efficiency that are rarely investigated together in graph representation learning. In WIDEN, we propose a novel inductive, meta path-free message passing scheme that packs up heterogeneous node features with their associated edges from both low- and high-order neighbor nodes. To further improve the training efficiency, we innovatively present an active downsampling strategy that drops unimportant neighbor nodes to facilitate faster information propagation. Experiments on three real-world heterogeneous graphs have further validated the efficacy of WIDEN on both transductive and inductive node representation learning, as well as the superior training efficiency against state-of-the-art baselines.
Music source separation (MSS) is the task of separating a music piece into individual sources, such as vocals and accompaniment. Recently, neural network based methods have been applied to address the MSS problem, and can be categorized into spectrogram and time-domain based methods. However, there is a lack of research of using complementary information of spectrogram and time-domain inputs for MSS. In this article, we propose a CatNet framework that concatenates a UNet separation branch using spectrogram as input and a WavUNet separation branch using time-domain waveform as input for MSS. We propose an end-to-end and fully differentiable system that incorporate spectrogram calculation into CatNet. In addition, we propose a novel mix-audio data augmentation method that randomly mix audio segments from the same source as augmented audio segments for training. Our proposed CatNet MSS system achieves a state-of-the-art vocals separation source distortion ratio (SDR) of 7.54 dB, outperforming MMDenseNet of 6.57 dB evaluated on the MUSDB18 dataset.
Multi-source unsupervised domain adaptation~(MSDA) aims at adapting models trained on multiple labeled source domains to an unlabeled target domain. In this paper, we propose a novel multi-source domain adaptation framework based on collaborative learning for semantic segmentation. Firstly, a simple image translation method is introduced to align the pixel value distribution to reduce the gap between source domains and target domain to some extent. Then, to fully exploit the essential semantic information across source domains, we propose a collaborative learning method for domain adaptation without seeing any data from target domain. In addition, similar to the setting of unsupervised domain adaptation, unlabeled target domain data is leveraged to further improve the performance of domain adaptation. This is achieved by additionally constraining the outputs of multiple adaptation models with pseudo labels online generated by an ensembled model. Extensive experiments and ablation studies are conducted on the widely-used domain adaptation benchmark datasets in semantic segmentation. Our proposed method achieves 59.0\% mIoU on the validation set of Cityscapes by training on the labeled Synscapes and GTA5 datasets and unlabeled training set of Cityscapes. It significantly outperforms all previous state-of-the-arts single-source and multi-source unsupervised domain adaptation methods.
Many matrix completion methods assume that the data follows the uniform distribution. To address the limitation of this assumption, Chen et al. \cite{Chen20152999} propose to recover the matrix where the data follows the specific biased distribution. Unfortunately, in most real-world applications, the recovery of a data matrix appears to be incomplete, and perhaps even corrupted information. This paper considers the recovery of a low-rank matrix, where some observed entries are sampled in a \emph{biased distribution} suitably dependent on \emph{leverage scores} of a matrix, and some observed entries are uniformly corrupted. Our theoretical findings show that we can provably recover an unknown $n\times n$ matrix of rank $r$ from just about $O(nr\log^2 n)$ entries even when the few observed entries are corrupted with a small amount of noisy information. Empirical studies verify our theoretical results.