In this study, we define interaction components of different orders between two input variables based on game theory. We further prove that interaction components of different orders satisfy several desirable properties.
Fine-tuning pre-trained models have achieved impressive performance on standard natural language processing benchmarks. However, the resultant model generalizability remains poorly understood. We do not know, for example, how excellent performance can lead to the perfection of generalization models. In this study, we analyze a fine-tuned BERT model from different perspectives using relation extraction. We also characterize the differences in generalization techniques according to our proposed improvements. From empirical experimentation, we find that BERT suffers a bottleneck in terms of robustness by way of randomizations, adversarial and counterfactual tests, and biases (i.e., selection and semantic). These findings highlight opportunities for future improvements. Our open-sourced testbed DiagnoseRE is available in https://github.com/zjunlp/DiagnoseRE/.
Knowledge Graph Embedding (KGE) is a popular method for KG reasoning and usually a higher dimensional one ensures better reasoning capability. However, high-dimensional KGEs pose huge challenges to storage and computing resources and are not suitable for resource-limited or time-constrained applications, for which faster and cheaper reasoning is necessary. To address this problem, we propose DistilE, a knowledge distillation method to build low-dimensional student KGE from pre-trained high-dimensional teacher KGE. We take the original KGE loss as hard label loss and design specific soft label loss for different KGEs in DistilE. We also propose a two-stage distillation approach to make the student and teacher adapt to each other and further improve the reasoning capability of the student. Our DistilE is general enough to be applied to various KGEs. Experimental results of link prediction show that our method successfully distills a good student which performs better than a same dimensional one directly trained, and sometimes even better than the teacher, and it can achieve 2 times - 8 times embedding compression rate and more than 10 times faster inference speed than the teacher with a small performance loss. We also experimentally prove the effectiveness of our two-stage training proposal via ablation study.
This paper proposes a method to disentangle and quantify interactions among words that are encoded inside a DNN for natural language processing. We construct a tree to encode salient interactions extracted by the DNN. Six metrics are proposed to analyze properties of interactions between constituents in a sentence. The interaction is defined based on Shapley values of words, which are considered as an unbiased estimation of word contributions to the network prediction. Our method is used to quantify word interactions encoded inside the BERT, ELMo, LSTM, CNN, and Transformer networks. Experimental results have provided a new perspective to understand these DNNs, and have demonstrated the effectiveness of our method.
In order to prevent leaking input information from intermediate-layer features, this paper proposes a method to revise the traditional neural network into the rotation-equivariant neural network (RENN). Compared to the traditional neural network, the RENN uses d-ary vectors/tensors as features, in which each element is a d-ary number. These d-ary features can be rotated (analogous to the rotation of a d-dimensional vector) with a random angle as the encryption process. Input information is hidden in this target phase of d-ary features for attribute obfuscation. Even if attackers have obtained network parameters and intermediate-layer features, they cannot extract input information without knowing the target phase. Hence, the input privacy can be effectively protected by the RENN. Besides, the output accuracy of RENNs only degrades mildly compared to traditional neural networks, and the computational cost is significantly less than the homomorphic encryption.
This paper presents a method to interpret the success of knowledge distillation by quantifying and analyzing task-relevant and task-irrelevant visual concepts that are encoded in intermediate layers of a deep neural network (DNN). More specifically, three hypotheses are proposed as follows. 1. Knowledge distillation makes the DNN learn more visual concepts than learning from raw data. 2. Knowledge distillation ensures that the DNN is prone to learning various visual concepts simultaneously. Whereas, in the scenario of learning from raw data, the DNN learns visual concepts sequentially. 3. Knowledge distillation yields more stable optimization directions than learning from raw data. Accordingly, we design three types of mathematical metrics to evaluate feature representations of the DNN. In experiments, we diagnosed various DNNs, and above hypotheses were verified.
The goal of representation learning of knowledge graph is to encode both entities and relations into a low-dimensional embedding spaces. Many recent works have demonstrated the benefits of knowledge graph embedding on knowledge graph completion task, such as relation extraction. However, we observe that: 1) existing method just take direct relations between entities into consideration and fails to express high-order structural relationship between entities; 2) these methods just leverage relation triples of KGs while ignoring a large number of attribute triples that encoding rich semantic information. To overcome these limitations, this paper propose a novel knowledge graph embedding method, named KANE, which is inspired by the recent developments of graph convolutional networks (GCN). KANE can capture both high-order structural and attribute information of KGs in an efficient, explicit and unified manner under the graph convolutional networks framework. Empirical results on three datasets show that KANE significantly outperforms seven state-of-arts methods. Further analysis verify the efficiency of our method and the benefits brought by the attention mechanism.
Network embedding has been intensively studied in the literature and widely used in various applications, such as link prediction and node classification. While previous work focus on the design of new algorithms or are tailored for various problem settings, the discussion of initialization strategies in the learning process is often missed. In this work, we address this important issue of initialization for network embedding that could dramatically improve the performance of the algorithms on both effectiveness and efficiency. Specifically, we first exploit the graph partition technique that divides the graph into several disjoint subsets, and then construct an abstract graph based on the partitions. We obtain the initialization of the embedding for each node in the graph by computing the network embedding on the abstract graph, which is much smaller than the input graph, and then propagating the embedding among the nodes in the input graph. With extensive experiments on various datasets, we demonstrate that our initialization technique significantly improves the performance of the state-of-the-art algorithms on the evaluations of link prediction and node classification by up to 7.76% and 8.74% respectively. Besides, we show that the technique of initialization reduces the running time of the state-of-the-arts by at least 20%.
The main challenge of single image super resolution (SISR) is the recovery of high frequency details such as tiny textures. However, most of the state-of-the-art methods lack specific modules to identify high frequency areas, causing the output image to be blurred. We propose an attention-based approach to give a discrimination between texture areas and smooth areas. After the positions of high frequency details are located, high frequency compensation is carried out. This approach can incorporate with previously proposed SISR networks. By providing high frequency enhancement, better performance and visual effect are achieved. We also propose our own SISR network composed of DenseRes blocks. The block provides an effective way to combine the low level features and high level features. Extensive benchmark evaluation shows that our proposed method achieves significant improvement over the state-of-the-art works in SISR.