The LSTM network was proposed to overcome the difficulty in learning long-term dependence, and has made significant advancements in applications. With its success and drawbacks in mind, this paper raises the question - do RNN and LSTM have long memory? We answer it partially by proving that RNN and LSTM do not have long memory from a statistical perspective. A new definition for long memory networks is further introduced, and it requires the model weights to decay at a polynomial rate. To verify our theory, we convert RNN and LSTM into long memory networks by making a minimal modification, and their superiority is illustrated in modeling long-term dependence of various datasets.
相關內容
End-to-end approaches have drawn much attention recently for significantly simplifying the construction of an automatic speech recognition (ASR) system. RNN transducer (RNN-T) is one of the popular end-to-end methods. Previous studies have shown that RNN-T is difficult to train and a very complex training process is needed for a reasonable performance. In this paper, we explore RNN-T for a Chinese large vocabulary continuous speech recognition (LVCSR) task and aim to simplify the training process while maintaining performance. First, a new strategy of learning rate decay is proposed to accelerate the model convergence. Second, we find that adding convolutional layers at the beginning of the network and using ordered data can discard the pre-training process of the encoder without loss of performance. Besides, we design experiments to find a balance among the usage of GPU memory, training circle and model performance. Finally, we achieve 16.9% character error rate (CER) on our test set which is 2% absolute improvement from a strong BLSTM CE system with language model trained on the same text corpus.
In this paper, we proposed a new deep learning based dense monocular SLAM method. Compared to existing methods, the proposed framework constructs a dense 3D model via a sparse to dense mapping using learned surface normals. With single view learned depth estimation as prior for monocular visual odometry, we obtain both accurate positioning and high quality depth reconstruction. The depth and normal are predicted by a single network trained in a tightly coupled manner.Experimental results show that our method significantly improves the performance of visual tracking and depth prediction in comparison to the state-of-the-art in deep monocular dense SLAM.
Although end-to-end neural text-to-speech (TTS) methods (such as Tacotron2) are proposed and achieve state-of-the-art performance, they still suffer from two problems: 1) low efficiency during training and inference; 2) hard to model long dependency using current recurrent neural networks (RNNs). Inspired by the success of Transformer network in neural machine translation (NMT), in this paper, we introduce and adapt the multi-head attention mechanism to replace the RNN structures and also the original attention mechanism in Tacotron2. With the help of multi-head self-attention, the hidden states in the encoder and decoder are constructed in parallel, which improves the training efficiency. Meanwhile, any two inputs at different times are connected directly by self-attention mechanism, which solves the long range dependency problem effectively. Using phoneme sequences as input, our Transformer TTS network generates mel spectrograms, followed by a WaveNet vocoder to output the final audio results. Experiments are conducted to test the efficiency and performance of our new network. For the efficiency, our Transformer TTS network can speed up the training about 4.25 times faster compared with Tacotron2. For the performance, rigorous human tests show that our proposed model achieves state-of-the-art performance (outperforms Tacotron2 with a gap of 0.048) and is very close to human quality (4.39 vs 4.44 in MOS).
In recent years, the sequence-to-sequence learning neural networks with attention mechanism have achieved great progress. However, there are still challenges, especially for Neural Machine Translation (NMT), such as lower translation quality on long sentences. In this paper, we present a hierarchical deep neural network architecture to improve the quality of long sentences translation. The proposed network embeds sequence-to-sequence neural networks into a two-level category hierarchy by following the coarse-to-fine paradigm. Long sentences are input by splitting them into shorter sequences, which can be well processed by the coarse category network as the long distance dependencies for short sentences is able to be handled by network based on sequence-to-sequence neural network. Then they are concatenated and corrected by the fine category network. The experiments shows that our method can achieve superior results with higher BLEU(Bilingual Evaluation Understudy) scores, lower perplexity and better performance in imitating expression style and words usage than the traditional networks.
During the last years, there has been a lot of interest in achieving some kind of complex reasoning using deep neural networks. To do that, models like Memory Networks (MemNNs) have combined external memory storages and attention mechanisms. These architectures, however, lack of more complex reasoning mechanisms that could allow, for instance, relational reasoning. Relation Networks (RNs), on the other hand, have shown outstanding results in relational reasoning tasks. Unfortunately, their computational cost grows quadratically with the number of memories, something prohibitive for larger problems. To solve these issues, we introduce the Working Memory Network, a MemNN architecture with a novel working memory storage and reasoning module. Our model retains the relational reasoning abilities of the RN while reducing its computational complexity from quadratic to linear. We tested our model on the text QA dataset bAbI and the visual QA dataset NLVR. In the jointly trained bAbI-10k, we set a new state-of-the-art, achieving a mean error of less than 0.5%. Moreover, a simple ensemble of two of our models solves all 20 tasks in the joint version of the benchmark.
Template-matching methods for visual tracking have gained popularity recently due to their comparable performance and fast speed. However, they lack effective ways to adapt to changes in the target object's appearance, making their tracking accuracy still far from state-of-the-art. In this paper, we propose a dynamic memory network to adapt the template to the target's appearance variations during tracking. An LSTM is used as a memory controller, where the input is the search feature map and the outputs are the control signals for the reading and writing process of the memory block. As the location of the target is at first unknown in the search feature map, an attention mechanism is applied to concentrate the LSTM input on the potential target. To prevent aggressive model adaptivity, we apply gated residual template learning to control the amount of retrieved memory that is used to combine with the initial template. Unlike tracking-by-detection methods where the object's information is maintained by the weight parameters of neural networks, which requires expensive online fine-tuning to be adaptable, our tracker runs completely feed-forward and adapts to the target's appearance changes by updating the external memory. Moreover, the capacity of our model is not determined by the network size as with other trackers -- the capacity can be easily enlarged as the memory requirements of a task increase, which is favorable for memorizing long-term object information. Extensive experiments on OTB and VOT demonstrates that our tracker MemTrack performs favorably against state-of-the-art tracking methods while retaining real-time speed of 50 fps.
Model compression is significant for the wide adoption of Recurrent Neural Networks (RNNs) in both user devices possessing limited resources and business clusters requiring quick responses to large-scale service requests. This work aims to learn structurally-sparse Long Short-Term Memory (LSTM) by reducing the sizes of basic structures within LSTM units, including input updates, gates, hidden states, cell states and outputs. Independently reducing the sizes of basic structures can result in inconsistent dimensions among them, and consequently, end up with invalid LSTM units. To overcome the problem, we propose Intrinsic Sparse Structures (ISS) in LSTMs. Removing a component of ISS will simultaneously decrease the sizes of all basic structures by one and thereby always maintain the dimension consistency. By learning ISS within LSTM units, the obtained LSTMs remain regular while having much smaller basic structures. Based on group Lasso regularization, our method achieves 10.59x speedup without losing any perplexity of a language modeling of Penn TreeBank dataset. It is also successfully evaluated through a compact model with only 2.69M weights for machine Question Answering of SQuAD dataset. Our approach is successfully extended to non- LSTM RNNs, like Recurrent Highway Networks (RHNs). Our source code is publicly available at //github.com/wenwei202/iss-rnns
To solve the text-based question and answering task that requires relational reasoning, it is necessary to memorize a large amount of information and find out the question relevant information from the memory. Most approaches were based on external memory and four components proposed by Memory Network. The distinctive component among them was the way of finding the necessary information and it contributes to the performance. Recently, a simple but powerful neural network module for reasoning called Relation Network (RN) has been introduced. We analyzed RN from the view of Memory Network, and realized that its MLP component is able to reveal the complicate relation between question and object pair. Motivated from it, we introduce which uses MLP to find out relevant information on Memory Network architecture. It shows new state-of-the-art results in jointly trained bAbI-10k story-based question answering tasks and bAbI dialog-based question answering tasks.
In this short note, we present an extension of long short-term memory (LSTM) neural networks to using a depth gate to connect memory cells of adjacent layers. Doing so introduces a linear dependence between lower and upper layer recurrent units. Importantly, the linear dependence is gated through a gating function, which we call depth gate. This gate is a function of the lower layer memory cell, the input to and the past memory cell of this layer. We conducted experiments and verified that this new architecture of LSTMs was able to improve machine translation and language modeling performances.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191