Attention-based Encoder-decoder Models Research Articles

A framework-based transformer and knowledge distillation for interior style classification

Interior style classification is an interesting problem which has potential applications both commercial and academic communities. This task aims to devise interior design styles automatically. Thus, interior designers will explore customers’ tastes and then precisely provide suggestions for decor inspiration based on their preferences. Recently, Convolutional Neural Networks (CNNs) have been considered the de-facto standard in computer vision tasks. Therefore, several current works have tended to address interior style classification using CNN-based architectures. Moreover, transformer-based architectures and attention-based encoder–decoder models have been proven successfully in computer vision and natural language processing tasks. Sequentially, more studies have been arguing the efficiency of combining CNN-based architectures and transformer-based architectures for normal image classification problems. In this project, we focus on finding an architecture network that is suitable for the interior style classification problem. We propose a robustness method to address interior style design classification, named ISC-DeIT. The proposed method is based on Data-efficient image transformer architectures and knowledge distillation, which can be trained on small datasets effectively. Especially, a proposed additional module is plugged to leverage learning feature representations for improving predictive accuracy. Experiments were carried out on a new curated dataset with five interior styles including Art-Decor, Hitech, Indochina, Industrial, and Scandinavian. Empirical results of ISC- DeiT indicated that the ability of prediction for interior style classification of the proposed method has been increased significantly, compared with other state-of-the-art methods.

A deep reinforcement learning approach for solving the Traveling Salesman Problem with Drone

Reinforcement learning has recently shown promise in learning quality solutions in many combinatorial optimization problems. In particular, the attention-based encoder-decoder models show high effectiveness on various routing problems, including the Traveling Salesman Problem (TSP). Unfortunately, they perform poorly for the TSP with Drone (TSP-D), requiring routing a heterogeneous fleet of vehicles in coordination—a truck and a drone. In TSP-D, the two vehicles are moving in tandem and may need to wait at a node for the other vehicle to join. State-less attention-based decoder fails to make such coordination between vehicles. We propose a hybrid model that uses an attention encoder and a Long Short-Term Memory (LSTM) network decoder, in which the decoder’s hidden state can represent the sequence of actions made. We empirically demonstrate that such a hybrid model improves upon a purely attention-based model for both solution quality and computational efficiency. Our experiments on the min-max Capacitated Vehicle Routing Problem (mmCVRP) also confirm that the hybrid model is more suitable for the coordinated routing of multiple vehicles than the attention-based model. The proposed model demonstrates comparable results as the operations research baseline methods.

Alignment Knowledge Distillation for Online Streaming Attention-Based Speech Recognition

This article describes an efficient training method for online streaming attention-based encoder-decoder (AED) automatic speech recognition (ASR) systems. AED models have achieved competitive performance in offline scenarios by jointly optimizing all components. They have recently been extended to an online streaming framework via models such as monotonic chunkwise attention (MoChA). However, the elaborate attention calculation process is not robust for long-form speech utterances. Moreover, the sequence-level training objective and time-restricted streaming encoder cause a nonnegligible delay in token emission during inference. To address these problems, we propose CTC synchronous training (CTC-ST), in which CTC alignments are leveraged as a reference for token boundaries to enable a MoChA model to learn optimal monotonic input-output alignments. We formulate a purely end-to-end training objective to synchronize the boundaries of MoChA to those of CTC. The CTC model shares an encoder with the MoChA model to enhance the encoder representation. Moreover, the proposed method provides alignment information learned in the CTC branch to the attention-based decoder. Therefore, CTC-ST can be regarded as self-distillation of alignment knowledge from CTC to MoChA. Experimental evaluations on a variety of benchmark datasets show that the proposed method significantly reduces recognition errors and emission latency simultaneously, especially for long-form and noisy speech. We also compare CTC-ST with several methods that distill alignment knowledge from a hybrid ASR system and show that the CTC-ST can achieve a comparable tradeoff of accuracy and latency without relying on external alignment information. The best MoChA system shows performance comparable to that of RNN-transducer (RNN-T).

A CTC Alignment-Based Non-Autoregressive Transformer for End-to-End Automatic Speech Recognition

Recently, end-to-end models have been widely used in automatic speech recognition (ASR) systems. Two of the most representative approaches are connectionist temporal classification (CTC) and attention-based encoder-decoder (AED) models. Autoregressive transformers, variants of AED, adopt an autoregressive mechanism for token generation and thus are relatively slow during inference. In this paper, we present a comprehensive study of a CTC Alignment-based Single-Step Non-Autoregressive Transformer (CASS-NAT) for end-to-end ASR. In CASS-NAT, word embeddings in the autoregressive transformer (AT) are substituted with token-level acoustic embeddings (TAE) that are extracted from encoder outputs with the acoustical boundary information offered by the CTC alignment. TAE can be obtained in parallel, resulting in a parallel generation of output tokens. During training, Viterbi-alignment is used for TAE generation, and multiple training strategies are further explored to improve the word error rate (WER) performance. During inference, an error-based alignment sampling method is investigated in depth to reduce the alignment mismatch in the training and testing processes. Experimental results show that the CASS-NAT has a WER that is close to AT on various ASR tasks, while providing a ~24x inference speedup. With and without self-supervised learning, we achieve new state-of-the-art results for non-autoregressive models on several datasets. We also analyze the behavior of the CASS-NAT decoder to explain why it can perform similarly to AT. We find that TAEs have similar functionality to word embeddings for grammatical structures, which might indicate the possibility of learning some semantic information from TAEs without a language model.

Combining hybrid DNN-HMM ASR systems with attention-based models using lattice rescoring

The traditional hybrid deep neural network (DNN)–hidden Markov model (HMM) system and attention-based encoder–decoder (AED) model are both commonly used automatic speech recognition (ASR) approaches with distinct characteristics and advantages. While hybrid systems are per-frame-based and highly modularised to leverage external phonetic and linguistic knowledge, AED models operate on a per-label basis and jointly learn the acoustic and language information using a single model in an end-to-end trainable fashion. In this paper, we propose combining these two approaches in a two-pass rescoring framework. The first-pass uses hybrid ASR systems to facilitate streaming and controllable ASR, and the second-pass re-scores the N-best hypotheses or lattices produced by the first-pass hybrid DNN-HMM system with AED models. We also propose an improved algorithm for lattice rescoring with AED models. Experiments show the combined two-pass systems achieve competitive performance without using extra speech or text data on two standard ASR tasks. For the 80-hour AMI IHM dataset, the combined system has a 13.7% word error rate (WER) on the evaluation set and is up to a 29% relative WER reduction over the individual systems. For the 300-hour Switchboard dataset, the WERs of the combined system are 5.7% and 12.1% on Switchboard and CallHome subsets of Hub5’00, and 13.2% and 7.6% on Switchboard Cellular and Fisher subsets of RT03, and are up to a 33% relative reduction in WER over the individual systems.

Handwritten Mathematical Expression Recognition via Attention Aggregation Based Bi-directional Mutual Learning

Handwritten mathematical expression recognition aims to automatically generate LaTeX sequences from given images. Currently, attention-based encoder-decoder models are widely used in this task. They typically generate target sequences in a left-to-right (L2R) manner, leaving the right-to-left (R2L) contexts unexploited. In this paper, we propose an Attention aggregation based Bi-directional Mutual learning Network (ABM) which consists of one shared encoder and two parallel inverse decoders (L2R and R2L). The two decoders are enhanced via mutual distillation, which involves one-to-one knowledge transfer at each training step, making full use of the complementary information from two inverse directions. Moreover, in order to deal with mathematical symbols in diverse scales, an Attention Aggregation Module (AAM) is proposed to effectively integrate multi-scale coverage attentions. Notably, in the inference phase, given that the model already learns knowledge from two inverse directions, we only use the L2R branch for inference, keeping the original parameter size and inference speed. Extensive experiments demonstrate that our proposed approach achieves the recognition accuracy of 56.85 % on CROHME 2014, 52.92 % on CROHME 2016, and 53.96 % on CROHME 2019 without data augmentation and model ensembling, substantially outperforming the state-of-the-art methods. The source code is available in https://github.com/XH-B/ABM.

Hybrid Autoregressive and Non-Autoregressive Transformer Models for Speech Recognition

The autoregressive (AR) models, such as attention-based encoder-decoder models and RNN-Transducer, have achieved great success in speech recognition. They predict the output sequence conditioned on the previous tokens and acoustic encoded states, which is inefficient on GPUs. The non-autoregressive (NAR) models can get rid of the temporal dependency between the output tokens and predict the entire output tokens in at least one step. However, the NAR model still faces two major problems. On the one hand, there is still a great gap in performance between the NAR models and the advanced AR models. On the other hand, it's difficult for most of the NAR models to train and converge. To address these two problems, we propose a new model named the two-step non-autoregressive transformer(TSNAT), which improves the performance and accelerating the convergence of the NAR model by learning prior knowledge from a parameters-sharing AR model. Furthermore, we introduce the two-stage method into the inference process, which improves the model performance greatly. All the experiments are conducted on a public Chinese mandarin dataset ASIEHLL-1. The results show that the TSNAT can achieve a competitive performance with the AR model and outperform many complicated NAR models.

Combining Frame-Synchronous and Label-Synchronous Systems for Speech Recognition

Commonly used automatic speech recognition (ASR) systems can be classified into frame-synchronous and label-synchronous categories, based on whether the speech is decoded on a per-frame or per-label basis. Frame-synchronous systems, such as traditional hidden Markov model systems, can easily incorporate existing knowledge and can support streaming ASR applications. Label-synchronous systems, based on attention-based encoder-decoder models, can jointly learn the acoustic and language information with a single model, which can be regarded as audio-grounded language models. In this paper, we propose rescoring the N-best hypotheses or lattices produced by a first-pass frame-synchronous system with a label-synchronous system in a second-pass. By exploiting the complementary modelling of the different approaches, the combined two-pass systems achieve competitive performance without using any extra speech or text data on two standard ASR tasks. For the 80-hour AMI IHM dataset, the combined system has a 13.7% word error rate (WER) on the evaluation set, which is up to a 29% relative WER reduction over the individual systems. For the 300-hour Switchboard dataset, the WERs of the combined system are 5.7% and 12.1% on Switchboard and CallHome subsets of Hub5'00, and 13.2% and 7.6% on Switchboard Cellular and Fisher subsets of RT03, up to a 33% relative reduction in WER over the individual systems.

Self-Supervised Pre-Training for Attention-Based Encoder-Decoder ASR Model

End-to-end (E2E) models, including the attention-based encoder-decoder (AED) models, have achieved promising performance on the automatic speech recognition (ASR) task. However, the supervised training process of the E2E model needs a large amount of speech-text paired data. In contrast, self-supervised pre-training can pre-train the model on the unlabeled data and then fine-tune it on the limited labeled data to realize better performance. Most of the previous self-supervised pre-training methods focus on learning hidden representations from speech but ignore how to utilize the unpaired text. As a result, previous works often pre-train an acoustic encoder and then fine-tune it as a classification based ASR model, such as Connectionist Temporal Classification (CTC) based model, rather than an AED model. In this paper, we propose a self-supervised pre-training method for the AED model (SP-AED). The SP-AED method contains acoustic pre-training for the encoder, linguistic pre-training for the decoder, and an adaptive combination fine-tuning for the whole system. We first design a linguistic pre-training method for decoder by utilizing the text-only data. The decoder will be pre-trained as a noise-condition language model to learn the prior distribution of the text. Then, we pre-train the AED encoder with the wav2vec2.0 method with some modifications. Finally, we combine the pre-trained encoder and decoder and fine-tune them on the limited labeled data. We design an adaptive combination method during fine-tuning by modifying the decoder's input and output to prevent catastrophic forgetting. Experiments prove that compared with the random initialized models, the SP-AED pre-trained models can realize up to 17% relative improvement. And with similar model size or computational cost, we can get comparable results to other classification-based models on both English and Chinese corpus.

Integrating Knowledge Into End-to-End Speech Recognition From External Text-Only Data

Attention-based encoder-decoder (AED) models have achieved promising performance in speech recognition. However, because of the end-to-end training, an AED model is usually trained with speech-text paired data. It is challenging to incorporate external text-only data into AED models. Another issue of the AED model is that it does not use the right context of a text token while predicting the token. To alleviate the above two issues, we propose a unified method called LST (Learn Spelling from Teachers) to integrate knowledge into an AED model from the external text-only data and leverage the whole context in a sentence. The method is divided into two stages. First, in the representation stage, a language model is trained on the text. It can be seen as that the knowledge in the text is compressed into the LM. Then, at the transferring stage, the knowledge is transferred to the AED model via teacher-student learning. To further use the whole context of the text sentence, we propose an LM called causal cloze completer (COR), which estimates the probability of a token, given both the left context and the right context of it. Therefore, with LST training, the AED model can leverage the whole context in the sentence. Different from fusion based methods, which use LM during decoding, the proposed method does not increase any extra complexity at the inference stage. We conduct experiments on two scales of public Chinese datasets AISHELL-1 and AISHELL-2. The experimental results demonstrate the effectiveness of leveraging external text-only data and the whole context in a sentence with our proposed method, compared with baseline hybrid systems and AED model based systems.

Gated Recurrent Context: Softmax-Free Attention for Online Encoder-Decoder Speech Recognition

Recently, attention-based encoder-decoder (AED) models have shown state-of-the-art performance in automatic speech recognition (ASR). As the original AED models with global attentions are not capable of online inference, various online attention schemes have been developed to reduce ASR latency for better user experience. However, a common limitation of the conventional softmax-based online attention approaches is that they introduce an additional hyperparameter related to the length of the attention window, requiring multiple trials of model training for tuning the hyperparameter. In order to deal with this problem, we propose a novel softmax-free attention method and its modified formulation for online attention, which does not need any additional hyperparameter at the training phase. Through a number of ASR experiments, we demonstrate the tradeoff between the latency and performance of the proposed online attention technique can be controlled by merely adjusting a threshold at the test phase. Furthermore, the proposed methods showed competitive performance to the conventional global and online attentions in terms of word-error-rates (WERs).

Differentiable Neural Architecture Search for High-Dimensional Time Series Forecasting

This study applies neural architecture search (NAS) techniques to the modeling of high-dimensional time series data such as multi-variate stock indices. It is known that traditional NAS method applies fully connected directed acyclic graph (DAG) for searching cell structures that requires high computational cost and cannot include the two-input operations such as the Hadamard product. To address the drawback of the DAG backbone, a novel two-input backbone cell architecture for recurrent neural networks is proposed, in which each candidate operation is also carried out with two inputs. Instead of using DAG, we simplify the backbone by considering the prior knowledge as an effective backbone such as preserving identity mappings. The cell structures will be incorporated in different types of model architectures including stacked long short-term memory (LSTM), gated recurrent unit (GRU) and attention-based encoder-decoder models. The experimental results on BRICS, G7 and G20 indices indicate that models with recurrent neural network (RNN) cells searched by the proposed backbone structure can significantly outperform baseline models including autoregressive integrated moving average model (ARIMA), vector autoregression (VAR), and stacked LSTM/GRUs. For neural architecture search, the proposed backbone is shown to be more effective compared to the classic differentiable architecture search (DARTS) in both uni-variate and multi-variate time series prediction tasks. Further analysis demonstrates that the pruned cells of the proposed backbone usually contains the Hadamard product introduced as a two-input operation, while the number of parameters involved in these pruned cells is on the same order with the baseline cells.

Fast End-to-End Speech Recognition Via Non-Autoregressive Models and Cross-Modal Knowledge Transferring From BERT

Attention-based encoder-decoder (AED) models have achieved promising performance in speech recognition. However, because the decoder predicts text tokens (such as characters or words) in an autoregressive manner, it is difficult for an AED model to predict all tokens in parallel. This makes the inference speed relatively slow. We believe that because the encoder already captures the whole speech utterance, which has the token-level relationship implicitly, we can predict a token without explicitly autoregressive language modeling. When the prediction of a token does not rely on other tokens, the parallel prediction of all tokens in the sequence is realizable. Based on this idea, we propose a non-autoregressive speech recognition model called LASO (Listen Attentively, and Spell Once). The model consists of an encoder, a decoder, and a position dependent summarizer (PDS). The three modules are based on basic attention blocks. The encoder extracts high-level representations from the speech. The PDS uses positional encodings corresponding to tokens to convert the acoustic representations into token-level representations. The decoder further captures token-level relationships with the self-attention mechanism. At last, the probability distribution on the vocabulary is computed for each token position. Therefore, speech recognition is re-formulated as a position-wise classification problem. Further, we propose a cross-modal transfer learning method to refine semantics from a large-scale pre-trained language model BERT for improving the performance.

Deep Learning-based Roman-Urdu to Urdu Transliteration

Attention-based encoder-decoder models have superseded conventional techniques due to their unmatched performance on many neural machine translation problems. Usually, the encoders and decoders are two recurrent neural networks where the decoder is directed to focus on relevant parts of the source language using attention mechanism. This data-driven approach leads to generic and scalable solutions with no reliance on manual hand-crafted features. To the best of our knowledge, none of the modern machine translation approaches has been applied to address the research problem of Urdu machine transliteration. Ours is the first attempt to apply the deep neural network-based encoder-decoder using attention mechanism to address the aforementioned problem using Roman-Urdu and Urdu parallel corpus. To this end, we present (i) the first ever Roman-Urdu to Urdu parallel corpus of 1.1 million sentences, (ii) three state of the art encoder-decoder models, and (iii) a detailed empirical analysis of these three models on the Roman-Urdu to Urdu parallel corpus. Overall, attention-based model gives state-of-the-art performance with the benchmark of 70 BLEU score. Our qualitative experimental evaluation shows that our models generate coherent transliterations which are grammatically and logically correct.

Segment boundary detection directed attention for online end-to-end speech recognition

Attention-based encoder-decoder models have recently shown competitive performance for automatic speech recognition (ASR) compared to conventional ASR systems. However, how to employ attention models for online speech recognition still needs to be explored. Different from conventional attention models wherein the soft alignment is obtained by a pass over the entire input sequence, attention models for online recognition must learn online alignment to attend part of input sequence monotonically when generating output symbols. Based on the fact that every output symbol is corresponding to a segment of input sequence, we propose a new attention mechanism for learning online alignment by decomposing the conventional alignment into two parts: segmentation—segment boundary detection with hard decision—and segment-directed attention—information aggregation within the segment with soft attention. The boundary detection is conducted along the time axis from left to right, and a decision is made for each input frame about whether it is a segment boundary or not. When a boundary is detected, the decoder generates an output symbol by attending the inputs within the corresponding segment. With the proposed attention mechanism, online speech recognition can be realized. The experimental results on TIMIT and WSJ dataset show that our proposed attention mechanism achieves comparable online performance with state-of-the-art models.