We present a generic and flexible Reinforcement Learning (RL) based meta-learning framework for the problem of few-shot learning. During training, it learns the best optimization algorithm to produce a learner (ranker/classifier, etc) by exploiting stable patterns in loss surfaces. Our method implicitly estimates the gradients of a scaled loss function while retaining the general properties intact for parameter updates. Besides providing improved performance on few-shot tasks, our framework could be easily extended to do network architecture search. We further propose a novel dual encoder, affinity-score based decoder topology that achieves additional improvements to performance. Experiments on an internal dataset, MQ2007, and AwA2 show our approach outperforms existing alternative approaches by 21%, 8%, and 4% respectively on accuracy and NDCG metrics. On Mini-ImageNet dataset our approach achieves comparable results with Prototypical Networks. Empirical evaluations demonstrate that our approach provides a unified and effective framework.
Fingerprint image denoising is a very important step in fingerprint identification. to improve the denoising effect of fingerprint image,we have designs a fingerprint denoising algorithm based on deep encoder-decoder network,which encoder subnet to learn the fingerprint features of noisy images.the decoder subnet reconstructs the original fingerprint image based on the features to achieve denoising, while using the dilated convolution in the network to increase the receptor field without increasing the complexity and improve the network inference speed. In addition, feature fusion at different levels of the network is achieved through the introduction of residual learning, which further restores the detailed features of the fingerprint and improves the denoising effect. Finally, the experimental results show that the algorithm enables better recovery of edge, line and curve features in fingerprint images, with better visual effects and higher peak signal-to-noise ratio (PSNR) compared to other methods.
This paper investigates how to effectively incorporate a pre-trained masked language model (MLM), such as BERT, into an encoder-decoder (EncDec) model for grammatical error correction (GEC). The answer to this question is not as straightforward as one might expect because the previous common methods for incorporating a MLM into an EncDec model have potential drawbacks when applied to GEC. For example, the distribution of the inputs to a GEC model can be considerably different (erroneous, clumsy, etc.) from that of the corpora used for pre-training MLMs; however, this issue is not addressed in the previous methods. Our experiments show that our proposed method, where we first fine-tune a MLM with a given GEC corpus and then use the output of the fine-tuned MLM as additional features in the GEC model, maximizes the benefit of the MLM. The best-performing model achieves state-of-the-art performances on the BEA-2019 and CoNLL-2014 benchmarks. Our code is publicly available at: https://github.com/kanekomasahiro/bert-gec.
Polysynthetic languages have exceptionally large and sparse vocabularies, thanks to the number of morpheme slots and combinations in a word. This complexity, together with a general scarcity of written data, poses a challenge to the development of natural language technologies. To address this challenge, we offer linguistically-informed approaches for bootstrapping a neural morphological analyzer, and demonstrate its application to Kunwinjku, a polysynthetic Australian language. We generate data from a finite state transducer to train an encoder-decoder model. We improve the model by "hallucinating" missing linguistic structure into the training data, and by resampling from a Zipf distribution to simulate a more natural distribution of morphemes. The best model accounts for all instances of reduplication in the test set and achieves an accuracy of 94.7% overall, a 10 percentage point improvement over the FST baseline. This process demonstrates the feasibility of bootstrapping a neural morph analyzer from minimal resources.
Semantic role labeling (SRL) is an NLP task involving the assignment of predicate arguments to types, called semantic roles. Though research on SRL has primarily focused on verbal predicates and many resources available for SRL provide annotations only for verbs, semantic relations are often triggered by other linguistic constructions, e.g., nominalizations. In this work, we investigate a transfer scenario where we assume role-annotated data for the source verbal domain but only unlabeled data for the target nominal domain. Our key assumption, enabling the transfer between the two domains, is that selectional preferences of a role (i.e., preferences or constraints on the admissible arguments) do not strongly depend on whether the relation is triggered by a verb or a noun. For example, the same set of arguments can fill the Acquirer role for the verbal predicate `acquire' and its nominal form `acquisition'. We approach the transfer task from the variational autoencoding perspective. The labeler serves as an encoder (predicting role labels given a sentence), whereas selectional preferences are captured in the decoder component (generating arguments for the predicting roles). Nominal roles are not labeled in the training data, and the learning objective instead pushes the labeler to assign roles predictive of the arguments. Sharing the decoder parameters across the domains encourages consistency between labels predicted for both domains and facilitates the transfer. The method substantially outperforms baselines, such as unsupervised and `direct transfer' methods, on the English CoNLL-2009 dataset.
In this paper, we present an Attention-based Identity Preserving Generative Adversarial Network (AIP-GAN) to overcome the identity leakage problem from a source image to a generated face image, an issue that is encountered in a cross-subject facial expression transfer and synthesis process. Our key insight is that the identity preserving network should be able to disentangle and compose shape, appearance, and expression information for efficient facial expression transfer and synthesis. Specifically, the expression encoder of our AIP-GAN disentangles the expression information from the input source image by predicting its facial landmarks using our supervised spatial and channel-wise attention module. Similarly, the disentangled expression-agnostic identity features are extracted from the input target image by inferring its combined intrinsic-shape and appearance image employing our self-supervised spatial and channel-wise attention mod-ule. To leverage the expression and identity information encoded by the intermediate layers of both of our encoders, we combine these features with the features learned by the intermediate layers of our decoder using a cross-encoder bilinear pooling operation. Experimental results show the promising performance of our AIP-GAN based technique.
It is a challenging task to extract the best of both worlds by combining the spatial characteristics of a visible image and the spectral content of an infrared image. In this work, we propose a spatially constrained adversarial autoencoder that extracts deep features from the infrared and visible images to obtain a more exhaustive and global representation. In this paper, we propose a residual autoencoder architecture, regularised by a residual adversarial network, to generate a more realistic fused image. The residual module serves as primary building for the encoder, decoder and adversarial network, as an add on the symmetric skip connections perform the functionality of embedding the spatial characteristics directly from the initial layers of encoder structure to the decoder part of the network. The spectral information in the infrared image is incorporated by adding the feature maps over several layers in the encoder part of the fusion structure, which makes inference on both the visual and infrared images separately. In order to efficiently optimize the parameters of the network, we propose an adversarial regulariser network which would perform supervised learning on the fused image and the original visual image.
There has been recent success in pre-training on monolingual data and fine-tuning on Machine Translation (MT), but it remains unclear how to best leverage a pre-trained model for a given MT task. This paper investigates the benefits and drawbacks of freezing parameters, and adding new ones, when fine-tuning a pre-trained model on MT. We focus on 1) Fine-tuning a model trained only on English monolingual data, BART. 2) Fine-tuning a model trained on monolingual data from 25 languages, mBART. For BART we get the best performance by freezing most of the model parameters, and adding extra positional embeddings. For mBART we match or outperform the performance of naive fine-tuning for most language pairs with the encoder, and most of the decoder, frozen. The encoder-decoder attention parameters are most important to fine-tune. When constraining ourselves to an out-of-domain training set for Vietnamese to English we see the largest improvements over the fine-tuning baseline.
Despite its original goal to jointly learn to align and translate, prior researches suggest that Transformer captures poor word alignments through its attention mechanism. In this paper, we show that attention weights DO capture accurate word alignments and propose two novel word alignment induction methods Shift-Att and Shift-AET. The main idea is to induce alignments at the step when the to-be-aligned target token is the decoder input rather than the decoder output as in previous work. Shift-Att is an interpretation method that induces alignments from the attention weights of Transformer and does not require parameter update or architecture change. Shift-AET extracts alignments from an additional alignment module which is tightly integrated into Transformer and trained in isolation with supervision from symmetrized Shift-Att alignments. Experiments on three publicly available datasets demonstrate that both methods perform better than their corresponding neural baselines and Shift-AET significantly outperforms GIZA++ by 1.4-4.8 AER points.
Image memes have become a widespread tool used by people for interacting and exchanging ideas over social media, blogs, and open messengers. This work proposes to treat automatic image meme generation as a translation process, and further present an end to end neural and probabilistic approach to generate an image-based meme for any given sentence using an encoder-decoder architecture. For a given input sentence, an image meme is generated by combining a meme template image and a text caption where the meme template image is selected from a set of popular candidates using a selection module, and the meme caption is generated by an encoder-decoder model. An encoder is used to map the selected meme template and the input sentence into a meme embedding and a decoder is used to decode the meme caption from the meme embedding. The generated natural language meme caption is conditioned on the input sentence and the selected meme template. The model learns the dependencies between the meme captions and the meme template images and generates new memes using the learned dependencies. The quality of the generated captions and the generated memes is evaluated through both automated and human evaluation. An experiment is designed to score how well the generated memes can represent the tweets from Twitter conversations. Experiments on Twitter data show the efficacy of the model in generating memes for sentences in online social interaction.
In this paper, we propose a neural machine translation system for Wolof, a low-resource Niger-Congo language. First we gathered a parallel corpus of 70000 aligned French-Wolof sentences. Then we developped a baseline LSTM based encoder-decoder architecture which was further extended to bidirectional LSTMs with attention mechanisms. Our models are trained on a limited amount of parallel French-Wolof data of approximately 35000 parallel sentences. Experimental results on French-Wolof translation tasks show that our approach produces promising translations in extremely low-resource conditions. The best model was able to achieve a good performance of 47% BLEU score.
Machine reading comprehension methods that generate answers by referring to multiple passages for a question have gained much attention in AI and NLP communities. The current methods, however, do not investigate the relationships among multiple passages in the answer generation process, even though topics correlated among the passages may be answer candidates. Our method, called neural answer Generation through Unified Memories over Multiple Passages (GUM-MP), solves this problem as follows. First, it determines which tokens in the passages are matched to the question. In particular, it investigates matches between tokens in positive passages, which are assigned to the question, and those in negative passages, which are not related to the question. Next, it determines which tokens in the passage are matched to other passages assigned to the same question and at the same time it investigates the topics in which they are matched. Finally, it encodes the token sequences with the above two matching results into unified memories in the passage encoders and learns the answer sequence by using an encoder-decoder with a multiple-pointer-generator mechanism. As a result, GUM-MP can generate answers by pointing to important tokens present across passages. Evaluations indicate that GUM-MP generates much more accurate results than the current models do.
The celebrated Seq2Seq technique and its numerous variants achieve excellent performance on many tasks such as neural machine translation, semantic parsing, and math word problem solving. However, these models either only consider input objects as sequences while ignoring the important structural information for encoding, or they simply treat output objects as sequence outputs instead of structural objects for decoding. In this paper, we present a novel Graph-to-Tree Neural Networks, namely Graph2Tree consisting of a graph encoder and a hierarchical tree decoder, that encodes an augmented graph-structured input and decodes a tree-structured output. In particular, we investigated our model for solving two problems, neural semantic parsing and math word problem. Our extensive experiments demonstrate that our Graph2Tree model outperforms or matches the performance of other state-of-the-art models on these tasks.
Clustering algorithms have wide applications and play an important role in data analysis fields including time series data analysis. However, in time series analysis, most of the algorithms used signal shape features or the initial value of hidden variable of a neural network. Little has been discussed on the methods based on the generative model of the time series. In this paper, we propose a new clustering algorithm focusing on the generative process of the signal with a recurrent neural network and the variational Bayes method. Our experiments show that the proposed algorithm not only has a robustness against for phase shift, amplitude and signal length variations but also provide a flexible clustering based on the property of the variational Bayes method.
Attention-based sequence-to-sequence (seq2seq) speech synthesis has achieved extraordinary performance. But a studio-quality corpus with manual transcription is necessary to train such seq2seq systems. In this paper, we propose an approach to build high-quality and stable seq2seq based speech synthesis system using challenging found data, where training speech contains noisy interferences (acoustic noise) and texts are imperfect speech recognition transcripts (textual noise). To deal with text-side noise, we propose a VQVAE based heuristic method to compensate erroneous linguistic feature with phonetic information learned directly from speech. As for the speech-side noise, we propose to learn a noise-independent feature in the auto-regressive decoder through adversarial training and data augmentation, which does not need an extra speech enhancement model. Experiments show the effectiveness of the proposed approach in dealing with text-side and speech-side noise. Surpassing the denoising approach based on a state-of-the-art speech enhancement model, our system built on noisy found data can synthesize clean and high-quality speech with MOS close to the system built on the clean counterpart.
Residual network (ResNet) and densely connected network (DenseNet) have significantly improved the training efficiency and performance of deep convolutional neural networks (DCNNs) mainly for object classification tasks. In this paper, we propose an efficient network architecture by considering advantages of both networks. The proposed method is integrated into an encoder-decoder DCNN model for medical image segmentation. Our method adds additional skip connections compared to ResNet but uses significantly fewer model parameters than DenseNet. We evaluate the proposed method on a public dataset (ISIC 2018 grand-challenge) for skin lesion segmentation and a local brain MRI dataset. In comparison with ResNet-based, DenseNet-based and attention network (AttnNet) based methods within the same encoder-decoder network structure, our method achieves significantly higher segmentation accuracy with fewer number of model parameters than DenseNet and AttnNet. The code is available on GitHub (GitHub link: https://github.com/MinaJf/DRU-net).
Numerous deep learning architectures have been developed to accommodate the diversity of time series datasets across different domains. In this article, we survey common encoder and decoder designs used in both one-step-ahead and multi-horizon time series forecasting -- describing how temporal information is incorporated into predictions by each model. Next, we highlight recent developments in hybrid deep learning models, which combine well-studied statistical models with neural network components to improve pure methods in either category. Lastly, we outline some ways in which deep learning can also facilitate decision support with time series data.
In this paper, we propose a Multi-Scale Boosted Dehazing Network with Dense Feature Fusion based on the U-Net architecture. The proposed method is designed based on two principles, boosting and error feedback, and we show that they are suitable for the dehazing problem. By incorporating the Strengthen-Operate-Subtract boosting strategy in the decoder of the proposed model, we develop a simple yet effective boosted decoder to progressively restore the haze-free image. To address the issue of preserving spatial information in the U-Net architecture, we design a dense feature fusion module using the back-projection feedback scheme. We show that the dense feature fusion module can simultaneously remedy the missing spatial information from high-resolution features and exploit the non-adjacent features. Extensive evaluations demonstrate that the proposed model performs favorably against the state-of-the-art approaches on the benchmark datasets as well as real-world hazy images.
In neural network-based models for natural language processing (NLP), the largest part of the parameters often consists of word embeddings. Conventional models prepare a large embedding matrix whose size depends on the vocabulary size. Therefore, storing these models in memory and disk storage is costly. In this study, to reduce the total number of parameters, the embeddings for all words are represented by transforming a shared embedding. The proposed method, ALONE (all word embeddings from one), constructs the embedding of a word by modifying the shared embedding with a filter vector, which is word-specific but non-trainable. Then, we input the constructed embedding into a feed-forward neural network to increase its expressiveness. Naively, the filter vectors occupy the same memory size as the conventional embedding matrix, which depends on the vocabulary size. To solve this issue, we also introduce a memory-efficient filter construction approach. We indicate our ALONE can be used as word representation sufficiently through an experiment on the reconstruction of pre-trained word embeddings. In addition, we also conduct experiments on NLP application tasks: machine translation and summarization. We combined ALONE with the current state-of-the-art encoder-decoder model, the Transformer, and achieved comparable scores on WMT 2014 English-to-German translation and DUC 2004 very short summarization with less parameters.
Designing effective neural networks is fundamentally important in deep multimodal learning. Most existing works focus on a single task and design neural architectures manually, which are highly task-specific and hard to generalize to different tasks. In this paper, we devise a generalized deep multimodal neural architecture search (MMnas) framework for various multimodal learning tasks. Given multimodal input, we first define a set of primitive operations, and then construct a deep encoder-decoder based unified backbone, where each encoder or decoder block corresponds to an operation searched from a predefined operation pool. On top of the unified backbone, we attach task-specific heads to tackle different multimodal learning tasks. By using a gradient-based NAS algorithm, the optimal architectures for different tasks are learned efficiently. Extensive ablation studies, comprehensive analysis, and comparative experimental results show that the obtained MMnasNet significantly outperforms existing state-of-the-art approaches across three multimodal learning tasks (over five datasets), including visual question answering, image-text matching, and visual grounding.