3D convolutional neural networks have revealed superior performance in processing volumetric data such as video and medical imaging. However, the competitive performance by leveraging 3D networks results in huge computational costs, which are far beyond that of 2D networks. In this paper, we propose a novel Hilbert curve-based cross-dimensionality distillation approach that facilitates the knowledge of 3D networks to improve the performance of 2D networks. The proposed Hilbert Distillation (HD) method preserves the structural information via the Hilbert curve, which maps high-dimensional (>=2) representations to one-dimensional continuous space-filling curves. Since the distilled 2D networks are supervised by the curves converted from dimensionally heterogeneous 3D features, the 2D networks are given an informative view in terms of learning structural information embedded in well-trained high-dimensional representations. We further propose a Variable-length Hilbert Distillation (VHD) method to dynamically shorten the walking stride of the Hilbert curve in activation feature areas and lengthen the stride in context feature areas, forcing the 2D networks to pay more attention to learning from activation features. The proposed algorithm outperforms the current state-of-the-art distillation techniques adapted to cross-dimensionality distillation on two classification tasks. Moreover, the distilled 2D networks by the proposed method achieve competitive performance with the original 3D networks, indicating the lightweight distilled 2D networks could potentially be the substitution of cumbersome 3D networks in the real-world scenario.
Deep learning (DL)-based channel state information (CSI) feedback has shown promising potential to improve spectrum efficiency in massive MIMO systems. However, practical DL approaches require a sizeable CSI dataset for each scenario, and require large storage for multiple learned models. To overcome this costly barrier, we develop a solution for efficient training and deployment enhancement of DL-based CSI feedback by exploiting a lightweight translation model to cope with new CSI environments and by proposing novel dataset augmentation based on domain knowledge. Specifically, we first develop a deep unfolding CSI feedback network, SPTM2-ISTANet+, which employs spherical normalization to address the challenge of path loss variation. We also introduce an integration of a trainable measurement matrix and residual CSI recovery blocks within SPTM2-ISTANet+ to improve efficiency and accuracy. Using SPTM2-ISTANet+ as the anchor feedback model, we propose an efficient scenario-adaptive CSI feedback architecture. This new CSI-TransNet exploits a plug-in module for CSI translation consisting of a sparsity aligning function and lightweight DL module to reuse pretrained models in unseen environments. To work with small datasets, we propose a lightweight and general augmentation strategy based on domain knowledge. Test results demonstrate the efficacy and efficiency of the proposed solution for accurate CSI feedback given limited measurements for unseen CSI environments.
We consider linear coding for Gaussian two-way channels (GTWCs), in which each user generates the transmit symbols by linearly encoding both its message and the past received symbols (i.e., the feedback information) from the other user. In Gaussian one-way channels (GOWCs), Butman has proposed a well-developed model for linear encoding that encapsulates feedback information into transmit signals. However, such a model for GTWCs has not been well studied since the coupling of the encoding processes at the users in GTWCs renders the encoding design non-trivial and challenging. In this paper, we aim to fill this gap in the literature by extending the existing signal models in GOWCs to GTWCs. With our developed signal model for GTWCs, we formulate an optimization problem to jointly design the encoding/decoding schemes for both the users, aiming to minimize the weighted sum of their transmit powers under signal-to-noise ratio constraints. First, we derive an optimal form of the linear decoding schemes under any arbitrary encoding schemes employed at the users. Further, we provide new insights on the encoding design for GTWCs. In particular, we show that it is optimal that one of the users (i) does not transmit the feedback information to the other user at the last channel use, and (ii) transmits its message only over the last channel use. With these solution behaviors, we further simplify the problem and solve it via an iterative two-way optimization scheme. We numerically demonstrate that our proposed scheme for GTWCs achieves a better performance in terms of the transmit power compared to the existing counterparts, such as the non-feedback scheme and one-way optimization scheme.
This paper introduces a novel neural network -- the flow completion network (FCN) -- to infer the fluid dynamics, including the flow field and the force acting on the body, from the incomplete data based on Graph Convolution Attention Network. The FCN is composed of several graph convolution layers and spatial attention layers. It is designed to infer the velocity field and the vortex force contribution of the flow field when combined with the vortex force map (VFM) method. Compared with other neural networks adopted in fluid dynamics, the FCN is capable of dealing with both structured data and unstructured data. The performance of the proposed FCN is assessed by the computational fluid dynamics (CFD) data on the flow field around a circular cylinder. The force coefficients predicted by our model are validated against those obtained directly from CFD. Moreover, it is shown that our model effectively utilizes the existing flow field information and the gradient information simultaneously, giving a better performance than the traditional CNN-based and DNN-based models.
Multimodal pre-training breaks down the modality barriers and allows the individual modalities to be mutually augmented with information, resulting in significant advances in representation learning. However, graph modality, as a very general and important form of data, cannot be easily interacted with other modalities because of its non-regular nature. In this paper, we propose MMGA (Multimodal learning with Graph Alignment), a novel multimodal pre-training framework to incorporate information from graph (social network), image and text modalities on social media to enhance user representation learning. In MMGA, a multi-step graph alignment mechanism is proposed to add the self-supervision from graph modality to optimize the image and text encoders, while using the information from the image and text modalities to guide the graph encoder learning. We conduct experiments on the dataset crawled from Instagram. The experimental results show that MMGA works well on the dataset and improves the fans prediction task's performance. We release our dataset, the first social media multimodal dataset with graph, of 60,000 users labeled with specific topics based on 2 million posts to facilitate future research.
HDR(High Dynamic Range) video can reproduce realistic scenes more realistically, with a wider gamut and broader brightness range. HDR video resources are still scarce, and most videos are still stored in SDR (Standard Dynamic Range) format. Therefore, SDRTV-to-HDRTV Conversion (SDR video to HDR video) can significantly enhance the user's video viewing experience. Since the correlation between adjacent video frames is very high, the method utilizing the information of multiple frames can improve the quality of the converted HDRTV. Therefore, we propose a multi-frame fusion neural network \textbf{DSLNet} for SDRTV to HDRTV conversion. We first propose a dynamic spatial-temporal feature alignment module \textbf{DMFA}, which can align and fuse multi-frame. Then a novel spatial-temporal feature modulation module \textbf{STFM}, STFM extracts spatial-temporal information of adjacent frames for more accurate feature modulation. Finally, we design a quality enhancement module \textbf{LKQE} with large kernels, which can enhance the quality of generated HDR videos. To evaluate the performance of the proposed method, we construct a corresponding multi-frame dataset using HDR video of the HDR10 standard to conduct a comprehensive evaluation of different methods. The experimental results show that our method obtains state-of-the-art performance. The dataset and code will be released.
In real-world crowdsourcing annotation systems, due to differences in user knowledge and cultural backgrounds, as well as the high cost of acquiring annotation information, the supervision information we obtain might be insufficient and ambiguous. To mitigate the negative impacts, in this paper, we investigate a more general and broadly applicable learning problem, i.e. \emph{semi-supervised partial label learning}, and propose a novel method based on pseudo-labeling and contrastive learning. Following the key inventing principle, our method facilitate the partial label disambiguation process with unlabeled data and at the same time assign reliable pseudo-labels to weakly supervised examples. Specifically, our method learns from the ambiguous labeling information via partial cross-entropy loss. Meanwhile, high-accuracy pseudo-labels are generated for both partial and unlabeled examples through confidence-based thresholding and contrastive learning is performed in a hybrid unsupervised and supervised manner for more discriminative representations, while its supervision increases curriculumly. The two main components systematically work as a whole and reciprocate each other. In experiments, our method consistently outperforms all comparing methods by a significant margin and set up the first state-of-the-art performance for semi-supervised partial label learning on image benchmarks.
In this paper, we focus on video-to-text summarization and investigate how to best utilize multimodal information for summarizing long inputs (e.g., an hour-long TV show) into long outputs (e.g., a multi-sentence summary). We extend SummScreen (Chen et al., 2021), a dialogue summarization dataset consisting of transcripts of TV episodes with reference summaries, and create a multimodal variant by collecting corresponding full-length videos. We incorporate multimodal information into a pre-trained textual summarizer efficiently using adapter modules augmented with a hierarchical structure while tuning only 3.8\% of model parameters. Our experiments demonstrate that multimodal information offers superior performance over more memory-heavy and fully fine-tuned textual summarization methods.
Word or word-fragment based Language Models (LM) are typically preferred over character-based ones in many downstream applications. This may not be surprising as words seem more linguistically relevant units than characters. Words provide at least two kinds of relevant information: boundary information and meaningful units. However, word boundary information may be absent or unreliable in the case of speech input (word boundaries are not marked explicitly in the speech stream). Here, we systematically compare LSTMs as a function of the input unit (character, phoneme, word, word part), with or without gold boundary information. We probe linguistic knowledge in the networks at the lexical, syntactic and semantic levels using three speech-adapted black box NLP psycholinguistically-inpired benchmarks (pWUGGY, pBLIMP, pSIMI). We find that the absence of boundaries costs between 2\% and 28\% in relative performance depending on the task. We show that gold boundaries can be replaced by automatically found ones obtained with an unsupervised segmentation algorithm, and that even modest segmentation performance gives a gain in performance on two of the three tasks compared to basic character/phone based models without boundary information.
The problem of spike encoding of sound consists in transforming a sound waveform into spikes. It is of interest in many domains, including the development of audio-based spiking neural networks, where it is the first and most crucial stage of processing. Many algorithms have been proposed to perform spike encoding of sound. However, a systematic approach to quantitatively evaluate their performance is currently lacking. We propose the use of an information-theoretic framework to solve this problem. Specifically, we evaluate the coding efficiency of four spike encoding algorithms on two coding tasks that consist of coding the fundamental characteristics of sound: frequency and amplitude. The algorithms investigated are: Independent Spike Coding, Send-on-Delta coding, Ben's Spiker Algorithm, and Leaky Integrate-and-Fire coding. Using the tools of information theory, we estimate the information that the spikes carry on relevant aspects of an input stimulus. We find disparities in the coding efficiencies of the algorithms, where Leaky Integrate-and-Fire coding performs best. The information-theoretic analysis of their performance on these coding tasks provides insight on the encoding of richer and more complex sound stimuli.