Recurrent neural networks are a widely used class of neural architectures. They have, however, two shortcomings. First, they are often treated as black-box models and as such it is difficult to understand what exactly they learn as well as how they arrive at a particular prediction. Second, they tend to work poorly on sequences requiring long-term memorization, despite having this capacity in principle. We aim to address both shortcomings with a class of recurrent networks that use a stochastic state transition mechanism between cell applications. This mechanism, which we term state-regularization, makes RNNs transition between a finite set of learnable states. We evaluate state-regularized RNNs on (1) regular languages for the purpose of automata extraction; (2) non-regular languages such as balanced parentheses and palindromes where external memory is required; and (3) real-word sequence learning tasks for sentiment analysis, visual object recognition and text categorisation. We show that state-regularization (a) simplifies the extraction of finite state automata that display an RNN's state transition dynamic; (b) forces RNNs to operate more like automata with external memory and less like finite state machines, which potentiality leads to a more structural memory; (c) leads to better interpretability and explainability of RNNs by leveraging the probabilistic finite state transition mechanism over time steps.

Transformers are state-of-the-art in a wide range of NLP tasks and have also been applied to many real-world products. Understanding the reliability and certainty of transformer models is crucial for building trustable machine learning applications, e.g., medical diagnosis. Although many recent transformer extensions have been proposed, the study of the uncertainty estimation of transformer models is under-explored. In this work, we propose a novel way to enable transformers to have the capability of uncertainty estimation and, meanwhile, retain the original predictive performance. This is achieved by learning hierarchical stochastic self-attention that attends to values and a set of learnable centroids, respectively. Then new attention heads are formed with a mixture of sampled centroids using the Gumbel-Softmax trick. We theoretically show that the self-attention approximation by sampling from a Gumbel distribution is upper bounded. We empirically evaluate our model on two text classification tasks with both in-domain (ID) and out-of-domain (OOD) datasets. The experimental results demonstrate that our approach: (1) achieves the best predictive-uncertainty trade-off among compared methods; (2) exhibits very competitive (in most cases, better) predictive performance on ID datasets; (3) is on par with Monte Carlo dropout and ensemble methods in uncertainty estimation on OOD datasets.

Data sparsity and data imbalance are practical and challenging issues in cross-domain recommender systems (RSs). This paper addresses those problems by leveraging the concepts which derive from representation learning, adversarial learning, and transfer learning (particularly, domain adaptation). Although various transfer learning methods have shown promising performance in this context, our proposed novel method RecSys-DAN focuses on alleviating the cross-domain and within-domain data sparsity and data imbalance and learns transferable latent representations for users, items, and their interactions. Different from the existing approaches, the proposed method transfers the latent representations from a source domain to a target domain in an adversarial way. The mapping functions in the target domain are learned by playing a min-max game with an adversarial loss, aiming to generate domain indistinguishable representations for a discriminator. Four neural architectural instances of ResSys-DAN are proposed and explored. Empirical results on real-world Amazon data show that, even without using labeled data (i.e., ratings) in the target domain, RecSys-DAN achieves competitive performance as compared to the state-of-the-art supervised methods. More importantly, RecSys-DAN is highly flexible to both unimodal and multimodal scenarios, and thus it is more robust to the cold-start recommendation which is difficult for the previous methods.

Recurrent neural networks are a widely used class of neural architectures with two shortcomings. First, it is difficult to understand what exactly they learn. Second, they tend to work poorly on sequences requiring long-term memorization, despite having this capacity in principle. We aim to address both shortcomings with a class of recurrent networks that use a stochastic state transition mechanism between cell applications. This mechanism, which we term state-regularization, makes RNNs transition between a finite set of learnable states. We evaluate state-regularized RNNs on (1) regular languages for the purpose of automata extraction; (2) nonregular languages such as balanced parentheses, palindromes, and the copy task where external memory is required; and (3) real-word sequence learning tasks for sentiment analysis, visual object recognition, and language modeling. We show that state-regularization simplifies the extraction of finite state automata from the RNN’s state transition dynamics; forces RNNs to operate more like automata with external memory and less like finite state machines; and makes RNNs more interpretable.

Multimodal learning has shown promising performance in content-based recommendation due to the auxiliary user and item information of multiple modalities such as text and images.However, the problem of incomplete and missing modality is rarely explored and most existing methods fail in learning a recommendation model with missing or corrupted modalities.In this paper, we propose LRMM, a novel framework that mitigates not only the problem of missing modalities but also more generally the cold-start problem of recommender systems.We propose modality dropout (m-drop) and a multimodal sequential autoencoder (m-auto) to learn multimodal representations for complementing and imputing missing modalities.Extensive experiments on real-world Amazon data show that LRMM achieves state-of-the-art performance on rating prediction tasks.More importantly, LRMM is more robust to previous methods in alleviating data-sparsity and the cold-start problem.

Generating a novel and descriptive caption of an image is drawing increasing interests in computer vision, natural language processing, and multimedia communities. In this work, we propose an end-to-end trainable deep bidirectional LSTM (Bi-LSTM (Long Short-Term Memory)) model to address the problem. By combining a deep convolutional neural network (CNN) and two separate LSTM networks, our model is capable of learning long-term visual-language interactions by making use of history and future context information at high-level semantic space. We also explore deep multimodal bidirectional models, in which we increase the depth of nonlinearity transition in different ways to learn hierarchical visual-language embeddings. Data augmentation techniques such as multi-crop, multi-scale, and vertical mirror are proposed to prevent overfitting in training deep models. To understand how our models “translate” image to sentence, we visualize and qualitatively analyze the evolution of Bi-LSTM internal states over time. The effectiveness and generality of proposed models are evaluated on four benchmark datasets: Flickr8K, Flickr30K, MSCOCO, and Pascal1K datasets. We demonstrate that Bi-LSTM models achieve highly competitive performance on both caption generation and image-sentence retrieval even without integrating an additional mechanism (e.g., object detection, attention model). Our experiments also prove that multi-task learning is beneficial to increase model generality and gain performance. We also demonstrate the performance of transfer learning of the Bi-LSTM model significantly outperforms previous methods on the Pascal1K dataset.

A video contains rich perceptual information, such as visual appearance, motion and audio, which can be used for understanding the activities in videos. Recent works have shown the combination of appearance (spatial) and motion (temporal) clues can significantly improve human action recognition performance in videos. To further explore the multimodal representation of video in action recognition, We propose a framework to learn a multimodal representations from video appearance, motion as well as audio data. Convolutional Neural Networks (CNN) are trained for each modality respectively. For fusing multiple features extracted with CNNs, we propose to add a fusion layer on the top of CNNs to learn a joint video representation. In fusion phase, we investigate both early fusion and late fusion with Neural Network and Support Vector Machine. Compare to existing works, (1) our work measures the benefits of taking audio information into consideration and (2) implements sophisticated fusion methods. The effectiveness of proposed approach is evaluated on UCF101 and UCF101-50 (selected subset in which each video contains audio data) for action recognition. The experimental results show that different modalities are complementary to each other and multimodal representation can be beneficial for final prediction. Furthermore, proposed fusion approach achieves 85.1% accuracy in fusing spatial-temporal on UCF101 (split 1), which is very competitive to state-of-the-art works.

This work presents an end-to-end trainable deep bidirectional LSTM (Long-Short Term Memory) model for image captioning. Our model builds on a deep convolutional neural network (CNN) and two separate LSTM networks. It is capable of learning long term visual-language interactions by making use of history and future context information at high level semantic space. Two novel deep bidirectional variant models, in which we increase the depth of nonlinearity transition in different way, are proposed to learn hierarchical visual-language embeddings. Data augmentation techniques such as multi-crop, multi-scale and vertical mirror are proposed to prevent overfitting in training deep models. We visualize the evolution of bidirectional LSTM internal states over time and qualitatively analyze how our models "translate" image to sentence. Our proposed models are evaluated on caption generation and image-sentence retrieval tasks with three benchmark datasets: Flickr8K, Flickr30K and MSCOCO datasets. We demonstrate that bidirectional LSTM models achieve highly competitive performance to the state-of-the-art results on caption generation even without integrating additional mechanism (e.g. object detection, attention model etc.) and significantly outperform recent methods on retrieval task

We showcase a system for real-time video text recognition. The system is based on the standard workflow of text spotting system, which includes text detection and word recognition procedures. We apply deep neural networks in both procedures. In text localization stage, textual candidates are roughly captured by using a Maximally Stable Extremal Regions (MSERs) detector with high recall rate, false alarms are then eliminated by using Convolutional Neural Network (CNN ) verifier. For word recognition, we developed a skeleton based method for segmenting text region from its background, then a CNN based word recognizer is utilized for recognizing texts. Our current implementation demonstrates a real time performance for recognizing scene text by using a standard laptop with webcam. The word recognizer achieves competitive result to state-of-the-art methods by only using synthetical training data.

Cross-Modal mapping plays an essential role in multimedia information retrieval systems. However, most of existing work paid much attention on learning mapping functions but neglected the exploration of high-level semantic representation of modalities. Inspired by recent success of deep learning, in this paper, deep CNN (convolutional neural networks) features and topic features are utilized as visual and textual semantic representation respectively. To investigate the highly non-linear semantic correlation between image and text, we propose a regularized deep neural network(RE-DNN) for semantic mapping across modalities. By imposing intra-modal regularization as supervised pre-training, we finally learn a joint model which captures both intra-modal and inter-modal relationships. Our approach is superior to previous work in follows: (1) it explores high-level semantic correlations, (2) it requires little prior knowledge for model training, (3) it is able to tackle modality missing problem. Extensive experiments on benchmark Wikipedia dataset show RE-DNN outperforms the state-of-the-art approaches in cross-modal retrieval.

In this paper we showcase a system for real-time text detection and recognition. We apply deep features created by Convolutional Neural Networks (CNNs) for both text detection and word recognition task. For text detection we follow the common localization-verification scheme which already shown its excellent ability in numerous previous work. In text localization stage, textual regions are roughly detected by using a MSERs (Maximally Stable Extremal Regions) detector with high recall rate. False alarms are then eliminated by using a CNNs classifier, and remaining text regions are further grouped into words. In the word recognition stage, we developed an skeleton-based text binarization method for segmenting text from its background. A CNNs based recognizer is then applied for recognizing character. The initial experiments show the powerful ability of deep features for text classification comparing with commonly used visual features. Our current implementation demonstrates real-time performance for recognizing scene text by using a standard PC with webcam.