In the era of deep learning, word embeddings are essential when dealing with text tasks. However, storing and accessing these embeddings requires a large amount of space. This is not conducive to the deployment of these models on resource-limited devices. Combining the powerful compression capability of tensor products, we propose a word embedding compression method with morphological augmentation, Morphologically-enhanced Tensorized Embeddings (MorphTE). A word consists of one or more morphemes, the smallest units that bear meaning or have a grammatical function. MorphTE represents a word embedding as an entangled form of its morpheme vectors via the tensor product, which injects prior semantic and grammatical knowledge into the learning of embeddings. Furthermore, the dimensionality of the morpheme vector and the number of morphemes are much smaller than those of words, which greatly reduces the parameters of the word embeddings. We conduct experiments on tasks such as machine translation and question answering. Experimental results on four translation datasets of different languages show that MorphTE can compress word embedding parameters by about 20 times without performance loss and significantly outperforms related embedding compression methods.
Dynamic interaction graphs have been widely adopted to model the evolution of user-item interactions over time. There are two crucial factors when modelling user preferences for link prediction in dynamic interaction graphs: 1) collaborative relationship among users and 2) user personalized interaction patterns. Existing methods often implicitly consider these two factors together, which may lead to noisy user modelling when the two factors diverge. In addition, they usually require time-consuming parameter learning with back-propagation, which is prohibitive for real-time user preference modelling. To this end, this paper proposes FreeGEM, a parameter-free dynamic graph embedding method for link prediction. Firstly, to take advantage of the collaborative relationships, we propose an incremental graph embedding engine to obtain user/item embeddings, which is an Online-Monitor-Offline architecture consisting of an Online module to approximately embed users/items over time, a Monitor module to estimate the approximation error in real time and an Offline module to calibrate the user/item embeddings when the online approximation errors exceed a threshold. Meanwhile, we integrate attribute information into the model, which enables FreeGEM to better model users belonging to some under represented groups. Secondly, we design a personalized dynamic interaction pattern modeller, which combines dynamic time decay with attention mechanism to model user short-term interests. Experimental results on two link prediction tasks show that FreeGEM can outperform the state-of-the-art methods in accuracy while achieving over 36X improvement in efficiency. All code and datasets can be found in https://github.com/FudanCISL/FreeGEM.
Hand, the bearer of human productivity and intelligence, is receiving much attention due to the recent fever of digital twins. Among different hand morphable models, MANO has been widely used in vision and graphics community. However, MANO disregards textures and accessories, which largely limits its power to synthesize photorealistic hand data. In this paper, we extend MANO with Diverse Accessories and Rich Textures, namely DART. DART is composed of 50 daily 3D accessories which varies in appearance and shape, and 325 hand-crafted 2D texture maps covers different kinds of blemishes or make-ups. Unity GUI is also provided to generate synthetic hand data with user-defined settings, e.g., pose, camera, background, lighting, textures, and accessories. Finally, we release DARTset, which contains large-scale (800K), high-fidelity synthetic hand images, paired with perfect-aligned 3D labels. Experiments demonstrate its superiority in diversity. As a complement to existing hand datasets, DARTset boosts the generalization in both hand pose estimation and mesh recovery tasks. Raw ingredients (textures, accessories), Unity GUI, source code and DARTset are publicly available at dart2022.github.io
Currently, many face forgery detection methods aggregate spatial and frequency features to enhance the generalization ability and gain promising performance under the cross-dataset scenario. However, these methods only leverage one level frequency information which limits their expressive ability. To overcome these limitations, we propose a multi-scale wavelet transformer framework for face forgery detection. Specifically, to take full advantage of the multi-scale and multi-frequency wavelet representation, we gradually aggregate the multi-scale wavelet representation at different stages of the backbone network. To better fuse the frequency feature with the spatial features, frequency-based spatial attention is designed to guide the spatial feature extractor to concentrate more on forgery traces. Meanwhile, cross-modality attention is proposed to fuse the frequency features with the spatial features. These two attention modules are calculated through a unified transformer block for efficiency. A wide variety of experiments demonstrate that the proposed method is efficient and effective for both within and cross datasets.
We introduce GLM-130B, a bilingual (English and Chinese) pre-trained language model with 130 billion parameters. It is an attempt to open-source a 100B-scale model at least as good as GPT-3 and unveil how models of such a scale can be successfully pre-trained. Over the course of this effort, we face numerous unexpected technical and engineering challenges, particularly on loss spikes and disconvergence. In this paper, we introduce the training process of GLM-130B including its design choices, training strategies for both efficiency and stability, and engineering efforts. The resultant GLM-130B model offers significant outperformance over GPT-3 175B on a wide range of popular English benchmarks while the performance advantage is not observed in OPT-175B and BLOOM-176B. It also consistently and significantly outperforms ERNIE TITAN 3.0 260B -- the largest Chinese language model -- across related benchmarks. Finally, we leverage a unique scaling property of GLM-130B to reach INT4 quantization, without quantization aware training and with almost no performance loss, making it the first among 100B-scale models. More importantly, the property allows its effective inference on 4$\times$RTX 3090 (24G) or 8$\times$RTX 2080 Ti (11G) GPUs, the most ever affordable GPUs required for using 100B-scale models. The GLM-130B model weights are publicly accessible and its code, training logs, related toolkit, and lessons learned are open-sourced at https://github.com/THUDM/GLM-130B .
As an excellent aided communication tool, intelligent reflecting surface (IRS) can make a significant rate enhancement and coverage extension. In this paper, we present an investigation on beamforming in an IRS-aided directional modulation (DM) network. To fully explore the advantages of IRS, two beamforming methods with enhanced secrecy rate (SR) performance are proposed. The first method of maximizing secrecy rate (Max-SR) alternately optimizes confidential message (CM) beamforming vector, artificial noise (AN) beamforming vector and phase shift matrix. The first optimization vector is directly computed by the Rayleigh ratio and the last two are solved with generalized power iteration (GPI). This method is called Max-SR-GPI. To reduce the computational complexity, a new method of maximizing receive power with zero-forcing constraint (Max-RP-ZFC) of only reflecting CM and no AN is proposed. Simulation results show that the proposed two methods harvest about 30 percent rate gains over the cases of random-phase IRS and no IRS, and the proposed Max-SR-GPI performs slightly better than the Max-RP-ZFC in terms of SR, particularly in the small-large IRS.
Compared with the rapid development of single-frequency multi-polarization SAR image classification technology, there is less research on the land cover classification of multifrequency polarimetric SAR (MF-PolSAR) images. In addition, the current deep learning methods for MF-PolSAR classification are mainly based on convolutional neural networks (CNNs), only local spatiality is considered but the nonlocal relationship is ignored. Therefore, based on semantic interaction and nonlocal topological structure, this paper proposes the MF semantics and topology fusion network (MF-STFnet) to improve MF-PolSAR classification performance. In MF-STFnet, two kinds of classification are implemented for each band, semantic information-based (SIC) and topological property-based (TPC). They work collaboratively during MF-STFnet training, which can not only fully leverage the complementarity of bands, but also combine local and nonlocal spatial information to improve the discrimination between different categories. For SIC, the designed crossband interactive feature extraction module (CIFEM) is embedded to explicitly model the deep semantic correlation among bands, thereby leveraging the complementarity of bands to make ground objects more separable. For TPC, the graph sample and aggregate network (GraphSAGE) is employed to dynamically capture the representation of nonlocal topological relations between land cover categories. In this way, the robustness of classification can be further improved by combining nonlocal spatial information. Finally, an adaptive weighting fusion (AWF) strategy is proposed to merge inference from different bands, so as to make the MF joint classification decisions of SIC and TPC. The comparative experiments show that MF-STFnet can achieve more competitive classification performance than some state-of-the-art methods.
We devise neuro-dynamic state estimation (Neuro-DSE), a learning-based dynamic state estimation (DSE) algorithm for networked microgrids (NMs) under unknown subsystems. Our contributions include: 1) a data-driven Neuro-DSE algorithm for NMs DSE with partially unidentified dynamic models, which incorporates the neural-ordinary-differential-equations (ODE-Net) into Kalman filters; 2) a self-refining Neuro-DSE algorithm (Neuro-DSE+) which enables data-driven DSE under limited and noisy measurements by establishing an automatic filtering, augmenting and correcting framework; 3) a Neuro-KalmanNet-DSE algorithm which further integrates KalmanNet with Neuro-DSE to relieve the model mismatch of both neural- and physics-based dynamic models; and 4) an augmented Neuro-DSE for joint estimation of NMs states and unknown parameters (e.g., inertia). Extensive case studies demonstrate the efficacy of Neuro-DSE and its variants under different noise levels, control modes, power sources, observabilities and model knowledge, respectively.
Paper recommendation with user-generated keyword is to suggest papers that simultaneously meet user's interests and are relevant to the input keyword. This is a recommendation task with two queries, a.k.a. user ID and keyword. However, existing methods focus on recommendation according to one query, a.k.a. user ID, and are not applicable to solving this problem. In this paper, we propose a novel click-through rate (CTR) prediction model with heterogeneous graph neural network, called AMinerGNN, to recommend papers with two queries. Specifically, AMinerGNN constructs a heterogeneous graph to project user, paper, and keyword into the same embedding space by graph representation learning. To process two queries, a novel query attentive fusion layer is designed to recognize their importances dynamically and then fuse them as one query to build a unified and end-to-end recommender system. Experimental results on our proposed dataset and online A/B tests prove the superiority of AMinerGNN.
Current AI systems are designed to solve close-world problems with the assumption that the underlying world is remaining more or less the same. However, when dealing with real-world problems such assumptions can be invalid as sudden and unexpected changes can occur. To effectively deploy AI-powered systems in the real world, AI systems should be able to deal with open-world novelty quickly. Inevitably, dealing with open-world novelty raises an important question of novelty difficulty. Knowing whether one novelty is harder to deal with than another, can help researchers to train their systems systematically. In addition, it can also serve as a measurement of the performance of novelty robust AI systems. In this paper, we propose to define the novelty reaction difficulty as a relative difficulty of performing the known task after the introduction of the novelty. We propose a universal method that can be applied to approximate the difficulty. We present the approximations of the difficulty using our method and show how it aligns with the results of the evaluation of AI agents designed to deal with novelty.