There are thousands of languages on earth, but visual perception is shared among peoples. Existing multimodal neural machine translation (MNMT) methods achieve knowledge transfer by enforcing one encoder to learn shared representation across textual and visual modalities. However, the training and inference process heavily relies on well-aligned bilingual sentence - image triplets as input, which are often limited in quantity. In this paper, we hypothesize that visual imagination via synthesizing visual representation from source text could help the neural model map two languages with different symbols, thus helps the translation task. Our proposed end-to-end imagination-based machine translation model (ImagiT) first learns to generate semantic-consistent visual representation from source sentence, and then generate target sentence based on both text representation and imagined visual representation. Experiments demonstrate that our translation model benefits from visual imagination and significantly outperforms the text-only neural machine translation (NMT) baseline. We also conduct analyzing experiments, and the results show that imagination can help fill in missing information when performing the degradation strategy.
In this paper, we introduce Cross-modal Alignment with mixture experts Neural Network (CameNN) recommendation model for intral-city retail industry, which aims to provide fresh foods and groceries retailing within 5 hours delivery service arising for the outbreak of Coronavirus disease (COVID-19) pandemic around the world. We propose CameNN, which is a multi-task model with three tasks including Image to Text Alignment (ITA) task, Text to Image Alignment (TIA) task and CVR prediction task. We use pre-trained BERT to generate the text embedding and pre-trained InceptionV4 to generate image patch embedding (each image is split into small patches with the same pixels and treat each patch as an image token). Softmax gating networks follow to learn the weight of each transformer expert output and choose only a subset of experts conditioned on the input. Then transformer encoder is applied as the share-bottom layer to learn all input features' shared interaction. Next, mixture of transformer experts (MoE) layer is implemented to model different aspects of tasks. At top of the MoE layer, we deploy a transformer layer for each task as task tower to learn task-specific information. On the real word intra-city dataset, experiments demonstrate CameNN outperform baselines and achieve significant improvements on the image and text representation. In practice, we applied CameNN on CVR prediction in our intra-city recommender system which is one of the leading intra-city platforms operated in China.
Deep learning (DL)-based models have demonstrated good performance in medical image segmentation. However, the models trained on a known dataset often fail when performed on an unseen dataset collected from different centers, vendors and disease populations. In this work, we present a random style transfer network to tackle the domain generalization problem for multi-vendor and center cardiac image segmentation. Style transfer is used to generate training data with a wider distribution/ heterogeneity, namely domain augmentation. As the target domain could be unknown, we randomly generate a modality vector for the target modality in the style transfer stage, to simulate the domain shift for unknown domains. The model can be trained in a semi-supervised manner by simultaneously optimizing a supervised segmentation and an unsupervised style translation objective. Besides, the framework incorporates the spatial information and shape prior of the target by introducing two regularization terms. We evaluated the proposed framework on 40 subjects from the M\&Ms challenge2020, and obtained promising performance in the segmentation for data from unknown vendors and centers.
Non-autoregressive neural machine translation achieves remarkable inference acceleration compared to autoregressive models. However, current non-autoregressive models still fall behind their autoregressive counterparts in prediction accuracy. We attribute the accuracy gaps to two disadvantages of non-autoregressive models: a) learning simultaneous generation under the overly strong conditional independence assumption; b) lacking explicit target language modeling. In this paper, we propose Glancing Transformer (GLAT) to address the above disadvantages, which reduces the difficulty of learning simultaneous generation and introduces explicit target language modeling in the non-autoregressive setting at the same time. Experiments on several benchmarks demonstrate that our approach significantly improves the accuracy of non-autoregressive models without sacrificing any inference efficiency. In particular, GLAT achieves 30.91 BLEU on WMT 2014 German-English, which narrows the gap between autoregressive models and non-autoregressive models to less than 0.5 BLEU score.
Both image registration and label fusion in the multi-atlas segmentation (MAS) rely on the intensity similarity between target and atlas images. However, such similarity can be problematic when target and atlas images are acquired using different imaging protocols. High-level structure information can provide reliable similarity measurement for cross-modality images when cooperating with deep neural networks (DNNs). This work presents a new MAS framework for cross-modality images, where both image registration and label fusion are achieved by DNNs. For image registration, we propose a consistent registration network, which can jointly estimate forward and backward dense displacement fields (DDFs). Additionally, an invertible constraint is employed in the network to reduce the correspondence ambiguity of the estimated DDFs. For label fusion, we adapt a few-shot learning network to measure the similarity of atlas and target patches. Moreover, the network can be seamlessly integrated into the patch-based label fusion. The proposed framework is evaluated on the MM-WHS dataset of MICCAI 2017. Results show that the framework is effective in both cross-modality registration and segmentation.
Left atrial (LA) and atrial scar segmentation from late gadolinium enhanced magnetic resonance imaging (LGE MRI) is an important task in clinical practice. %, to guide ablation therapy and predict treatment results for atrial fibrillation (AF) patients. The automatic segmentation is however still challenging, due to the poor image quality, the various LA shapes, the thin wall, and the surrounding enhanced regions. Previous methods normally solved the two tasks independently and ignored the intrinsic spatial relationship between LA and scars. In this work, we develop a new framework, namely AtrialJSQnet, where LA segmentation, scar projection onto the LA surface, and scar quantification are performed simultaneously in an end-to-end style. We propose a mechanism of shape attention (SA) via an explicit surface projection, to utilize the inherent correlation between LA and LA scars. In specific, the SA scheme is embedded into a multi-task architecture to perform joint LA segmentation and scar quantification. Besides, a spatial encoding (SE) loss is introduced to incorporate continuous spatial information of the target, in order to reduce noisy patches in the predicted segmentation. We evaluated the proposed framework on 60 LGE MRIs from the MICCAI2018 LA challenge. Extensive experiments on a public dataset demonstrated the effect of the proposed AtrialJSQnet, which achieved competitive performance over the state-of-the-art. The relatedness between LA segmentation and scar quantification was explicitly explored and has shown significant performance improvements for both tasks. The code and results will be released publicly once the manuscript is accepted for publication via https://zmiclab.github.io/projects.html.
Efficiently building an adversarial attacker for natural language processing (NLP) tasks is a real challenge. Firstly, as the sentence space is discrete, it is difficult to make small perturbations along the direction of gradients. Secondly, the fluency of the generated examples cannot be guaranteed. In this paper, we propose MHA, which addresses both problems by performing Metropolis-Hastings sampling, whose proposal is designed with the guidance of gradients. Experiments on IMDB and SNLI show that our proposed MHA outperforms the baseline model on attacking capability. Adversarial training with MAH also leads to better robustness and performance.
It has been a common approach to pre-train a language model on a large corpus and fine-tune it on task-specific data. In practice, we observe that fine-tuning a pre-trained model on a small dataset may lead to over- and/or under-estimation problem. In this paper, we propose MC-Tailor, a novel method to alleviate the above issue in text generation tasks by truncating and transferring the probability mass from over-estimated regions to under-estimated ones. Experiments on a variety of text generation datasets show that MC-Tailor consistently and significantly outperforms the fine-tuning approach. Our code is available at this url.
Auto-regressive sequence generative models trained by Maximum Likelihood Estimation suffer the exposure bias problem in practical finite sample scenarios. The crux is that the number of training samples for Maximum Likelihood Estimation is usually limited and the input data distributions are different at training and inference stages. Many method shave been proposed to solve the above problem (Yu et al., 2017; Lu et al., 2018), which relies on sampling from the non-stationary model distribution and suffers from high variance or biased estimations. In this paper, we propose{\psi}-MLE, a new training scheme for auto-regressive sequence generative models, which is effective and stable when operating at large sample space encountered in text generation. We derive our algorithm from a new perspective of self-augmentation and introduce bias correction with density ratio estimation. Extensive experimental results on synthetic data and real-world text generation tasks demonstrate that our method stably outperforms Maximum Likelihood Estimation and other state-of-the-art sequence generative models in terms of both quality and diversity.