This paper investigates the fronthaul compression problem in a user-centric cloud radio access network, in which single-antenna users are served by a central processor (CP) cooperatively via a cluster of remote radio heads (RRHs). To satisfy the fronthaul capacity constraint, this paper proposes a transform-compress-forward scheme, which consists of well-designed transformation matrices and uniform quantizers. The transformation matrices perform dimension reduction in the uplink and dimension expansion in the downlink. To reduce the communication overhead for designing the transformation matrices, this paper further proposes a deep learning framework to first learn a suboptimal transformation matrix at each RRH based on the local channel state information (CSI), and then to refine it iteratively. To facilitate the refinement process, we propose an efficient signaling scheme that only requires the transmission of low-dimensional effective CSI and its gradient between the CP and RRH, and further, a meta-learning based gated recurrent unit network to reduce the number of signaling transmission rounds. For the sum-rate maximization problem, simulation results show that the proposed two-stage neural network can perform close to the fully cooperative global CSI based benchmark with significantly reduced communication overhead for both the uplink and the downlink. Moreover, using the first stage alone can already outperform the existing local CSI based benchmark.
This paper addresses the design of transmit precoder and receive combiner matrices to support $N_{\rm s}$ independent data streams over a time-division duplex (TDD) point-to-point massive multiple-input multiple-output (MIMO) channel with either a fully digital or a hybrid structure. The optimal precoder and combiner design amounts to finding the top-$N_{\rm s}$ singular vectors of the channel matrix, but the explicit estimation of the entire high-dimensional channel would require significant pilot overhead. Alternatively, prior works seek to find the precoding and combining matrices directly by exploiting channel reciprocity and by using the power iteration method, but its performance degrades in the low SNR regime. To tackle this challenging problem, this paper proposes a learning-based active sensing framework, where the transmitter and the receiver send pilots alternately using sensing beamformers that are actively designed as functions of previously received pilots. This is accomplished by using recurrent neural networks to summarize information from the historical observations into hidden state vectors, then using fully connected neural networks to learn the appropriate sensing beamformers in the next pilot stage and finally the transmit precoding and receive combiner matrices for data communications. Simulations demonstrate that the learning-based method outperforms existing approaches significantly and maintains superior performance even in low SNR regimes both in fully digital and hybrid MIMO scenarios.
While Large Language Models (LLMs) demonstrate impressive capabilities in text generation, we find that their ability has yet to be generalized to music, humanity's creative language. We introduce ChatMusician, an open-source LLM that integrates intrinsic musical abilities. It is based on continual pre-training and finetuning LLaMA2 on a text-compatible music representation, ABC notation, and the music is treated as a second language. ChatMusician can understand and generate music with a pure text tokenizer without any external multi-modal neural structures or tokenizers. Interestingly, endowing musical abilities does not harm language abilities, even achieving a slightly higher MMLU score. Our model is capable of composing well-structured, full-length music, conditioned on texts, chords, melodies, motifs, musical forms, etc, surpassing GPT-4 baseline. On our meticulously curated college-level music understanding benchmark, MusicTheoryBench, ChatMusician surpasses LLaMA2 and GPT-3.5 on zero-shot setting by a noticeable margin. Our work reveals that LLMs can be an excellent compressor for music, but there remains significant territory to be conquered. We release our 4B token music-language corpora MusicPile, the collected MusicTheoryBench, code, model and demo in GitHub.
Interactive Video Object Segmentation (iVOS) is a challenging task that requires real-time human-computer interaction. To improve the user experience, it is important to consider the user's input habits, segmentation quality, running time and memory consumption.However, existing methods compromise user experience with single input mode and slow running speed. Specifically, these methods only allow the user to interact with one single frame, which limits the expression of the user's intent.To overcome these limitations and better align with people's usage habits, we propose a framework that can accept multiple frames simultaneously and explore synergistic interaction across frames (SIAF). Concretely, we designed the Across-Frame Interaction Module that enables users to annotate different objects freely on multiple frames. The AFI module will migrate scribble information among multiple interactive frames and generate multi-frame masks. Additionally, we employ the id-queried mechanism to process multiple objects in batches. Furthermore, for a more efficient propagation and lightweight model, we design a truncated re-propagation strategy to replace the previous multi-round fusion module, which employs an across-round memory that stores important interaction information. Our SwinB-SIAF achieves new state-of-the-art performance on DAVIS 2017 (89.6%, J&F@60). Moreover, our R50-SIAF is more than 3 faster than the state-of-the-art competitor under challenging multi-object scenarios.
The emergence of novel the dummy data injection attack (DDIA) poses a severe threat to the secure and stable operation of power systems. These attacks are particularly perilous due to the minimal Euclidean spatial separation between the injected malicious data and legitimate data, rendering their precise detection challenging using conventional distance-based methods. Furthermore, existing research predominantly focuses on various machine learning techniques, often analyzing the temporal data sequences post-attack or relying solely on Euclidean spatial characteristics. Unfortunately, this approach tends to overlook the inherent topological correlations within the non-Euclidean spatial attributes of power grid data, consequently leading to diminished accuracy in attack localization. To address this issue, this study takes a comprehensive approach. Initially, it examines the underlying principles of these new DDIAs on power systems. Here, an intricate mathematical model of the DDIA is designed, accounting for incomplete topological knowledge and alternating current (AC) state estimation from an attacker's perspective. Subsequently, by integrating a priori knowledge of grid topology and considering the temporal correlations within measurement data and the topology-dependent attributes of the power grid, this study introduces temporal and spatial attention matrices. These matrices adaptively capture the spatio-temporal correlations within the attacks. Leveraging gated stacked causal convolution and graph wavelet sparse convolution, the study jointly extracts spatio-temporal DDIA features. Finally, the research proposes a DDIA localization method based on spatio-temporal graph neural networks. The accuracy and effectiveness of the DDIA model are rigorously demonstrated through comprehensive analytical cases.
Deep learning-based methods have made significant achievements in music source separation. However, obtaining good results while maintaining a low model complexity remains challenging in super wide-band music source separation. Previous works either overlook the differences in subbands or inadequately address the problem of information loss when generating subband features. In this paper, we propose SCNet, a novel frequency-domain network to explicitly split the spectrogram of the mixture into several subbands and introduce a sparsity-based encoder to model different frequency bands. We use a higher compression ratio on subbands with less information to improve the information density and focus on modeling subbands with more information. In this way, the separation performance can be significantly improved using lower computational consumption. Experiment results show that the proposed model achieves a signal to distortion ratio (SDR) of 9.0 dB on the MUSDB18-HQ dataset without using extra data, which outperforms state-of-the-art methods. Specifically, SCNet's CPU inference time is only 48% of HT Demucs, one of the previous state-of-the-art models.
In the rapidly evolving field of artificial intelligence, the creation and utilization of synthetic datasets have become increasingly significant. This report delves into the multifaceted aspects of synthetic data, particularly emphasizing the challenges and potential biases these datasets may harbor. It explores the methodologies behind synthetic data generation, spanning traditional statistical models to advanced deep learning techniques, and examines their applications across diverse domains. The report also critically addresses the ethical considerations and legal implications associated with synthetic datasets, highlighting the urgent need for mechanisms to ensure fairness, mitigate biases, and uphold ethical standards in AI development.
This paper addresses an uplink localization problem in which a base station (BS) aims to locate a remote user with the help of reconfigurable intelligent surfaces (RISs). We propose a strategy in which the user transmits pilots sequentially and the BS adaptively adjusts the sensing vectors, including the BS beamforming vector and multiple RIS reflection coefficients based on the observations already made, to eventually produce an estimated user position. This is a challenging active sensing problem for which finding an optimal solution involves searching through a complicated functional space whose dimension increases with the number of measurements. We show that the long short-term memory (LSTM) network can be used to exploit the latent temporal correlation between measurements to automatically construct scalable state vectors. Subsequently, the state vector is mapped to the sensing vectors for the next time frame via a deep neural network (DNN). A final DNN is used to map the state vector to the estimated user position. Numerical result illustrates the advantage of the active sensing design as compared to non-active sensing methods. The proposed solution produces interpretable results and is generalizable in the number of sensing stages. Remarkably, we show that a network with one BS and multiple RISs can outperform a comparable setting with multiple BSs.
Real-time multi-person pose estimation presents significant challenges in balancing speed and precision. While two-stage top-down methods slow down as the number of people in the image increases, existing one-stage methods often fail to simultaneously deliver high accuracy and real-time performance. This paper introduces RTMO, a one-stage pose estimation framework that seamlessly integrates coordinate classification by representing keypoints using dual 1-D heatmaps within the YOLO architecture, achieving accuracy comparable to top-down methods while maintaining high speed. We propose a dynamic coordinate classifier and a tailored loss function for heatmap learning, specifically designed to address the incompatibilities between coordinate classification and dense prediction models. RTMO outperforms state-of-the-art one-stage pose estimators, achieving 1.1% higher AP on COCO while operating about 9 times faster with the same backbone. Our largest model, RTMO-l, attains 74.8% AP on COCO val2017 and 141 FPS on a single V100 GPU, demonstrating its efficiency and accuracy. The code and models are available at https://github.com/open-mmlab/mmpose/tree/dev-1.x/projects/rtmo.