We study the learning of a matching model for dialogue response selection. Motivated by the recent finding that random negatives are often too trivial to train a reliable model, we propose a hierarchical curriculum learning (HCL) framework that consists of two complementary curricula: (1) corpus-level curriculum (CC); and (2) instance-level curriculum (IC). In CC, the model gradually increases its ability in finding the matching clues between the dialogue context and response. On the other hand, IC progressively strengthens the model's ability in identifying the mismatched information between the dialogue context and response. Empirical studies on two benchmark datasets with three state-of-the-art matching models demonstrate that the proposed HCL significantly improves the model performance across various evaluation metrics.
The width of a neural network matters since increasing the width will necessarily increase the model capacity. However, the performance of a network does not improve linearly with the width and soon gets saturated. To tackle this problem, we propose to increase the number of networks rather than purely scaling up the width. To prove it, one large network is divided into several small ones, and each of these small networks has a fraction of the original one's parameters. We then train these small networks together and make them see various views of the same data to learn different and complementary knowledge. During this co-training process, networks can also learn from each other. As a result, small networks can achieve better ensemble performance than the large one with few or no extra parameters or FLOPs. \emph{This reveals that the number of networks is a new dimension of effective model scaling, besides depth/width/resolution}. Small networks can also achieve faster inference speed than the large one by concurrent running on different devices. We validate the idea -- increasing the number of networks is a new dimension of effective model scaling -- with different network architectures on common benchmarks through extensive experiments. The code is available at \url{https://github.com/mzhaoshuai/SplitNet-Divide-and-Co-training}.
Motivated by the increasing popularity of intelligent editing assistant, we introduce and investigate the task of narrative incoherence detection: Given a (corrupted) long-form narrative, decide whether there exists some semantic discrepancy in the narrative flow. Specifically, we focus on the missing sentence and incoherent sentence detection. Despite its simple setup, this task is challenging as the model needs to understand and analyze a multi-sentence narrative text, and make decisions at the sentence level. As an initial step towards this task, we implement several baselines either directly analyzing the raw text (\textit{token-level}) or analyzing learned sentence representations (\textit{sentence-level}). We observe that while token-level modeling enjoys greater expressive power and hence better performance, sentence-level modeling possesses an advantage in efficiency and flexibility. With pre-training on large-scale data and cycle-consistent sentence embedding, our extended sentence-level model can achieve comparable detection accuracy to the token-level model. As a by-product, such a strategy enables simultaneous incoherence detection and infilling/modification suggestions.
The literature has witnessed the success of applying deep Transfer Learning (TL) algorithms to many NLP applications, yet it is not easy to build a simple and scalable TL toolkit for this purpose. To bridge this gap, the EasyTransfer platform is designed to make it easy to develop deep TL algorithms for NLP applications. It is built with rich API abstractions, a scalable architecture and comprehensive deep TL algorithms, to make the development of NLP applications easier. To be specific, the build-in data and model parallelism strategy shows to be 4x faster than the default distribution strategy of Tensorflow. EasyTransfer supports the mainstream pre-trained ModelZoo, including Pre-trained Language Models (PLMs) and multi-modality models. It also integrates various SOTA models for mainstream NLP applications in AppZoo, and supports mainstream TL algorithms as well. The toolkit is convenient for users to quickly start model training, evaluation, offline prediction, and online deployment. This system is currently deployed at Alibaba to support a variety of business scenarios, including item recommendation, personalized search, and conversational question answering. Extensive experiments on real-world datasets show that EasyTransfer is suitable for online production with cutting-edge performance. The source code of EasyTransfer is released at Github (https://github.com/alibaba/EasyTransfer).
Knowledge distillation aims at obtaining a small but effective deep model by transferring knowledge from a much larger one. The previous approaches try to reach this goal by simply "logit-supervised" information transferring between the teacher and student, which somehow can be subsequently decomposed as the transfer of normalized logits and $l^2$ norm. We argue that the norm of logits is actually interference, which damages the efficiency in the transfer process. To address this problem, we propose Spherical Knowledge Distillation (SKD). Specifically, we project the teacher and the student's logits into a unit sphere, and then we can efficiently perform knowledge distillation on the sphere. We verify our argument via theoretical analysis and ablation study. Extensive experiments have demonstrated the superiority and scalability of our method over the SOTAs.
Recently, there are emerging many stereo matching methods for autonomous driving based on unsupervised learning. Most of them take advantage of reconstruction losses to remove dependency on disparity groundtruth. Occlusion handling is a challenging problem in stereo matching, especially for unsupervised methods. Previous unsupervised methods failed to take full advantage of geometry properties in occlusion handling. In this paper, we introduce an effective way to detect occlusion regions and propose a novel unsupervised training strategy to deal with occlusion that only uses the predicted left disparity map, by making use of its geometry features in an iterative way. In the training process, we regard the predicted left disparity map as pseudo groundtruth and infer occluded regions using geometry features. The resulting occlusion mask is then used in either training, post-processing, or both of them as guidance. Experiments show that our method could deal with the occlusion problem effectively and significantly outperforms the other unsupervised methods for stereo matching. Moreover, our occlusion-aware strategies can be extended to the other stereo methods conveniently and improve their performances.
To deploy a pre-trained deep CNN on resource-constrained mobile devices, neural network pruning is often used to cut down the model's computational cost. For example, filter-level pruning (reducing the model's width) or layer-level pruning (reducing the model's depth) can both save computations with some sacrifice of accuracy. Besides, reducing the resolution of input images can also reach the same goal. Most previous methods focus on reducing one or two of these dimensions (i.e., depth, width, and image resolution) for acceleration. However, excessive reduction of any single dimension will lead to unacceptable accuracy loss, and we have to prune these three dimensions comprehensively to yield the best result. In this paper, a simple yet effective pruning framework is proposed to comprehensively consider these three dimensions. Our framework falls into two steps: 1) Determining the optimal depth (d*), width (w*), and image resolution (r) for the model. 2) Pruning the model in terms of (d*, w*, r*). Specifically, at the first step, we formulate model acceleration as an optimization problem. It takes depth (d), width (w) and image resolution (r) as variables and the model's accuracy as the optimization objective. Although it is hard to determine the expression of the objective function, approximating it with polynomials is still feasible, during which several properties of the objective function are utilized to ease and speedup the fitting process. Then the optimal d*, w* and r* are attained by maximizing the objective function with Lagrange multiplier theorem and KKT conditions. Extensive experiments are done on several popular architectures and datasets. The results show that we have outperformd the state-of-the-art pruning methods. The code will be published soon.
Adversarial training is currently the most powerful defense against adversarial examples. Previous empirical results suggest that adversarial training requires wider networks for better performances. Yet, it remains elusive how does neural network width affects model robustness. In this paper, we carefully examine the relation between network width and model robustness. We present an intriguing phenomenon that the increased network width may not help robustness. Specifically, we show that the model robustness is closely related to both natural accuracy and perturbation stability, a new metric proposed in our paper to characterize the model's stability under adversarial perturbations. While better natural accuracy can be achieved on wider neural networks, the perturbation stability actually becomes worse, leading to a potentially worse overall model robustness. To understand the origin of this phenomenon, we further relate the perturbation stability with the network's local Lipschitznesss. By leveraging recent results on neural tangent kernels, we show that larger network width naturally leads to worse perturbation stability. This suggests that to fully unleash the power of wide model architecture, practitioners should adopt a larger regularization parameter for training wider networks. Experiments on benchmark datasets confirm that this strategy could indeed alleviate the perturbation stability issue and improve the state-of-the-art robust models.