Decoding Complexity Research Articles

Balancing the encoder and decoder complexity in image compression for classification

This paper presents a study on the computational complexity of coding for machines, with a focus on image coding for classification. We first conduct a comprehensive set of experiments to analyze the size of the encoder (which encodes images to bitstreams), the size of the decoder (which decodes bitstreams and predicts class labels), and their impact on the rate–accuracy trade-off in compression for classification. Through empirical investigation, we demonstrate a complementary relationship between the encoder size and the decoder size, i.e., it is better to employ a large encoder with a small decoder and vice versa. Motivated by this relationship, we introduce a feature compression-based method for efficient image compression for classification. By compressing features at various layers of a neural network-based image classification model, our method achieves adjustable rate, accuracy, and encoder (or decoder) size using a single model. Experimental results on ImageNet classification show that our method achieves competitive results with existing methods while being much more flexible. The code will be made publicly available.

Physical layer security for confidential transmissions in frequency hopping-based downlink NOMA networks

Facing the exponential number of Internet of Things (IoT) devices and the scarcity of available resources, next-generation wireless networks have to meet very challenging performance targets in terms of providing massive access and ensuring higher spectral efficiency. In this vein, Non-Orthogonal Multiple Access (NOMA) has been widely recognized as one of the advantageous techniques to handle the proliferation of the IoT. Nevertheless, from a security standpoint, enabling a user to decode the signals of the other users, while using Successive Interference Cancellation (SIC), raises serious concerns regarding confidentiality and vulnerability to malicious attacks. Meanwhile, conventional security paradigms, such as upper-layer encryption and sophisticated authentication mechanisms, require high computational complexity and additional processing, which impose an overwhelming burden on energy-efficient IoT devices. Alternatively, Physical layer Security (PLS) has sparked a significant interest as a promising complement to cryptographic techniques. The key idea of PLS is to avail wireless communication properties to secure communications without adding complex encryption mechanisms at higher layers. In this paper, we propose a PLS approach based on a network coding technique to prevent eavesdroppers from decoding users’ information transmitted through a downlink-based NOMA system. This results in correlating the packets to be transmitted with each other, making the interception of a single packet useless. We demonstrate that the eavesdropper’s decoding complexity increases exponentially with the sequence length, making the task intractable for relatively long ones.

Shift Pruning-Based SCL Decoding for Polar Codes

In the context of the high-speed development of 5G communications, high-performance decoding schemes for polar codes are a hot spot in channel coding research. Shift pruning successive cancellation list (SP-SCL) decoding aims to recover the correct path by shift pruning in the extra SCL decoding. However, the current SP-SCL decoding is inflexible in determining the shift positions. In this paper, a flexible shift pruning SCL (FSP-SCL) decoding is proposed. Firstly, the reasons for movement and the eliminated states of the correct path are analyzed in detail using the path metric range (PMR), and on this basis, the validity of the method adopted in this paper for determining the shift priority of the information bits is verified. Secondly, the FSP-SCL decoding proposes two methods for determining the shift positions. One is the log-likelihood ratio (LLR) threshold method, which compares the LLR values of the eliminated paths on the shift bit with the corresponding LLR threshold to determine the shift positions. The other is the path distance method. It combines the minimum distance between the eliminated paths and the received vector with the path metrics to determine the shift positions. Both methods are more flexible and practical, as they can calculate the corresponding shift positions online based on a specific shift bit, avoiding the high complexity caused by the simulation method. Finally, this paper designs various experimental schemes to verify the decoding performance of the FSP-SCL. The experimental results show that in terms of error-correction performance, the LLR threshold-based FSP-SCL (FSPL (LLR threshold)) decoding, the path distance-based FSP-SCL (FSPL (path distance)) decoding and the existing SP-SCL decoding are roughly equal overall. In terms of decoding complexity, FSPL (LLR threshold) decoding is slightly better than FSPL (path distance) decoding, and the decoding complexity of both is lower than that of SP-SCL decoding, with the difference being more pronounced in the medium to high SNR regions.

Open-AI model Efficient Memory Reduce Management for the Large Language Models (LLMs) Serving with Paged Attention of sharing the KV Cashes

Abstract: High throughput serving of large language models (LLMs) requires batching sufficiently many requests at a time. However, existing systems struggle because the key-value cache (KV cache) memory for each request is huge and grows and shrinks dynamically. When managed inefficiently, this memory can be significantly wasted by fragmentation and redundant duplication, limiting the batch size. To address this problem we proposed a Paged Attention. An alternative algorithm inspired by the classical virtual memory and paging techniques in operating systems. An LLM serving system that achieves (1) near-zero waste in KV cache memory and (2) flexible sharing of KV cache within and across requests to further reduce memory usage. Our evaluations show that LLM improves the throughput of popular LLMs by 2-4× with the same level of latency compared to the state-of-the-art systems, such as Faster Transformer and Orca. The improvement is more pronounced with longer sequences, larger models, and more complex decoding algorithms

A multi-view ensemble machine learning approach for 3D modeling using geological and geophysical data

Geophysical data are often integrated into geological data for 3D modeling of underground spaces. However, the existing single-view approach means it is difficult to adequately fuse the valid information between the two types of data, and the complexity of lithological decoding and classification is high. To address this issue, a multi-view ensemble machine learning (ML) framework is proposed. Initially, the original dataset of lithology prediction is constructed by aligning geological and geophysical data with different spatial scales. Next, the dataset is divided into three datasets of structural strength, density, and moisture content according to the lithology properties of the geophysical data. The proposed framework is then used to capture the lithologic characteristics under different views to achieve the prediction of lithologic labels. In this process, a self-attentive mechanism is used to adaptively fuse the valid information under each view. To validate the proposed framework, it is applied to a project in Jiaxing, Zhejiang Province, China. Compared with existing ML methods, the proposed multi-view ensemble ML framework improves modeling accuracy and constructs models with low uncertainty. The framework can be extended to other multi-source data fusion tasks across geoscience domains.

In-Sensor Organic Electrochemical Transistor for the Multimode Neuromorphic Olfactory System.

The olfactory system is one of the six basic sensory nervous systems. Developing artificial olfactory systems is challenging due to the complexity of chemical information decoding and memory. Conventional chemical sensors can convert chemical signals into electric signals to decode gas information but they lack memory functions. Additional storage and processing units would significantly increase the complexity and power consumption of the devices, especially for portable and wearable devices. Here, an olfactory-inspired in-sensor organic electrochemical transistor (OI-OECT) is proposed, with the integrated functions of chemical information decoding, tunable memory level, and selectivity of vapor sensing. The ion-gel electrolyte endows the OI-OECT with the function of tunable memory levels and a low operating voltage. Typical synaptic behaviors, including inhibitory postsynaptic current and paired-pulse facilitations, are successfully achieved. Importantly, the gas memory level can be effectively modulated by the gate voltages (0 and -1 V), which realized the transformation of volatile and nonvolatile memory. Furthermore, benefiting from the recognition of multiple gases and ability to detect cumulative damage caused by gases, the OI-OECT is demonstrated for early warning system targeting leakage detection of two gases (NH3 and H2S). This work achieves the integrated functions of chemical gas information decode, tunable gas memory level, and selectivity of gas in a single device, which provides a promising pathway for the development of future artificial olfactory systems.

On the total energy consumption of scalable cache-aided multi-CPU cell-free massive MIMO systems

In conventional cell-free (CF) systems, the scalability and feasibility are hindered by the fact that a single central processing unit (CPU) manages all access points and the stringent synchronization requirement for coherent transmission (CT) across the entire network. Multi-CPU CF systems emerge as a scalable and feasible alternative, yet they still encounter the typical CF system challenge of high downlink transmit energy consumption from extensive fronthaul and backhaul transmissions. To tackle this issue, we propose a scalable downlink partial coherent transmission (PCT) strategy for multi-CPU CF systems to provide high spectral efficiency (SE), reduce transmission time and thereby lower energy consumption, while keeping decoding complexity manageable. Based on the PCT strategy, we introduce scalable cache-aided multi-CPU CF systems and develop the corresponding total energy consumption (TEC) model. A successive convex approximation algorithm is proposed to obtain the suboptimal cache placement to minimize the TEC. Simulation results indicate that the proposed PCT-based scalable cache-aided multi-CPU CF systems can significantly reduce the TEC and consistently maintain the TEC at a relatively low level across all Zipf parameters by leveraging the high SE of CT and the independent transmission of non-coherent transmission.

4 × 4 differential index modulation for optical orthogonal frequency division multiplexing.

In this Letter, we propose and demonstrate a 4 × 4 differential index modulation (DIM) scheme for optical orthogonal frequency division multiplexing (OOFDM) systems. The 4 × 4 DIM scheme avoids complex channel estimation by performing differential index modulation in the time-frequency domain, with the key aspect lying in the design of the time-frequency dispersion matrix that integrates the indices and constellation symbols. Moreover, we design a deep learning-based DIMFormer detector for the high decoding complexity problem of maximum likelihood (ML) detection. Experimental results show that the 4 × 4 DIM in OOFDM systems eliminates the need for complex channel estimation, and the loss of signal-to-noise ratio (SNR) is no more than 1 dB compared to conventional index modulation. The designed DIMFormer detector reduces the computational complexity by 38.98% as well as the time complexity by 99% compared to ML.

Dual model transfer learning to compensate for individual variability in brain-computer interface

Background and ObjectiveRecent advancements in brain-computer interface (BCI) technology have seen a significant shift towards incorporating complex decoding models such as deep neural networks (DNNs) to enhance performance. These models are particularly crucial for sophisticated tasks such as regression for decoding arbitrary movements. However, these BCI models trained and tested on individual data often face challenges with limited performance and generalizability across different subjects. This limitation is primarily due to a tremendous number of parameters of DNN models. Training complex models demands extensive datasets. Nevertheless, group data from many subjects may not produce sufficient decoding performance because of inherent variability in neural signals both across individuals and over time MethodsTo address these challenges, this study proposed a transfer learning approach that could effectively adapt to subject-specific variability in cortical regions. Our method involved training two separate movement decoding models: one on individual data and another on pooled group data. We then created a salience map for each cortical region from the individual model, which helped us identify the input's contribution variance across subjects. Based on the contribution variance, we combined individual and group models using a modified knowledge distillation framework. This approach allowed the group model to be universally applicable by assigning greater weights to input data, while the individual model was fine-tuned to focus on areas with significant individual variance ResultsOur combined model effectively encapsulated individual variability. We validated this approach with nine subjects performing arm-reaching tasks, with our method outperforming (mean correlation coefficient, r = 0.75) both individual (r = 0.70) and group models (r = 0.40) in decoding performance. In particular, there were notable improvements in cases where individual models showed low performances (e.g., r = 0.50 in the individual decoder to r = 0.61 in the proposed decoder) ConclusionsThese results not only demonstrate the potential of our method for robust BCI, but also underscore its ability to generalize individual data for broader applicability.

Now It Sounds Like You: Learning Personalized Vocabulary On Device

In recent years, Federated Learning (FL) has shown significant advancements in its ability to perform various natural language processing (NLP) tasks. This work focuses on applying personalized FL for on-device language modeling. Due to limitations of memory and latency, these models cannot support the complexity of sub-word tokenization or beam search decoding, resulting in the decision to deploy a closed-vocabulary language model. However, closed-vocabulary models are unable to handle out-of-vocabulary (OOV) words belonging to specific users. To address this issue, We propose a novel technique called "OOV expansion" that improves OOV coverage and increases model accuracy while minimizing the impact on memory and latency. This method introduces a personalized "OOV adapter" that effectively transfers knowledge from a central model and learns word embedding for personalized vocabulary. OOV expansion significantly outperforms standard FL personalization methods on a set of common FL benchmarks.

Investigation of IEEE 802.16e LDPC Code Application in PM-DQPSK System

With the development of the Internet and information technology, optical fiber communication systems need to meet people’s information demand for large capacity and high speed. High-order phase modulation and channel multiplexing can improve the capacity and data rate of optical fiber communication systems, but they also bring the problem of bit error. To improve the transmission quality and reliability of optical fiber communication systems, forward error correction (FEC) coding techniques are commonly used, which serve as the fundamental approach to enhance the quality and reliability of fiber optic communication systems, ensuring that the received data remain accurate and reliable. The FEC in optical fiber communication systems is divided into three generations. The first generation FEC is mainly hard decision codewords, represented as RS code. The second generation FEC is mainly cascaded code, which stands for interleaved cascaded code. The third generation of FEC mainly refers to soft decision codes, which are represented as low-density parity-check (LDPC) codes. As a kind of FEC, LDPC codes stand out as pivotal contributors in the field of optical communication and have gained remarkable attention due to exceptional error correction performance and low decoding complexity. Based on IEEE802.16e standard, LDPC code with specific code length and rate is compiled and simulated in MATLAB and VPItransmissionMaker 10.1 and successfully incorporated into polarization multiplexed differential quadrature phase shift keying (PM-DQPSK) coherent optical transmission system. The simulation results indicate that the bit error rate (BER) can be reduced to 10−3 when the optical signal-to-noise ratio (OSNR) reaches 14.2 dB, and the BER experiences a reduction by nearly three orders of magnitude when the OSNR is 17.2 dB. These findings underscore the efficacy of LDPC codes in significantly improving the performance of optical communication systems, particularly in scenarios demanding robust error correction capabilities. This study provides valuable, significant results regarding the potential of LDPC codes for enhancing the reliability of optical transmission in real-world applications.

Polar-coded channel polarization for reducing complexity of soft-decision forward error correction

To reduce the computational complexity of soft-decision (SD) forward error correction (FEC), we propose a polar coding method with a low-complexity successive cancellation decoder. Polar coding induces channel polarization in which two bit-channels with lower and higher reliabilities are polarized. Only the less-reliable bit-channels are protected by SD-FEC, whereas the more-reliable bit-channels are offloaded, reducing the complexity of SD-FEC decoding. The degradation of the bit error ratio (BER) performance can be suppressed by designing the polar encoder structures for the successive cancellation decoder. We numerically demonstrate that the proposed method manages to both reduce the computational complexity by half and suppress the BER performance degradation by less than 0.6 dB, compared with the conventional method using only the SD-FEC.

Reducing Complexity of SC-Based Flip Decoding of Polar Codes by Early-Stopping

Reducing Complexity of SC-Based Flip Decoding of Polar Codes by Early-Stopping

Concatenated Nanopore DNA Codes.

In nanopore sequencers, single-stranded DNA molecules (or k-mers) enter a small opening in a membrane called a nanopore and modulate the ionic current through the pore, producing a channel output in the form of a noisy piecewise constant signal. An important problem in DNA-based data storage is finding a set of k-mers, i.e. a DNA code, that is robust against noisy sample duplication introduced by nanopore sequencers. Good DNA codes should contain as many k-mers as possible that produce distinguishable current signals (squiggles) as measured by the sequencer. The dissimilarity between squiggles can be estimated using a bound on their pairwise error probability, which is used as a metric for code design. Unfortunately, code construction using the union bound is limited to small k's due to the difficulty of finding maximum cliques in large graphs. In this paper, we construct large codes by concatenating codewords from a base code, thereby packing more information in a single strand while retaining the storage efficiency of the base code. To facilitate decoding, we include a circumfix in the base code to reduce the effect of the nanopore channel memory. We show that the decoding complexity scales as [Formula: see text], where m is the number of concatenated k-mers. Simulations show that the base code error rate is stable as m increases.

BEV-MAE: Bird’s Eye View Masked Autoencoders for Point Cloud Pre-training in Autonomous Driving Scenarios

Existing LiDAR-based 3D object detection methods for autonomous driving scenarios mainly adopt the training-from-scratch paradigm. Unfortunately, this paradigm heavily relies on large-scale labeled data, whose collection can be expensive and time-consuming. Self-supervised pre-training is an effective and desirable way to alleviate this dependence on extensive annotated data. In this work, we present BEV-MAE, an efficient masked autoencoder pre-training framework for LiDAR-based 3D object detection in autonomous driving. Specifically, we propose a bird's eye view (BEV) guided masking strategy to guide the 3D encoder learning feature representation in a BEV perspective and avoid complex decoder design during pre-training. Furthermore, we introduce a learnable point token to maintain a consistent receptive field size of the 3D encoder with fine-tuning for masked point cloud inputs. Based on the property of outdoor point clouds in autonomous driving scenarios, i.e., the point clouds of distant objects are more sparse, we propose point density prediction to enable the 3D encoder to learn location information, which is essential for object detection. Experimental results show that BEV-MAE surpasses prior state-of-the-art self-supervised methods and achieves a favorably pre-training efficiency. Furthermore, based on TransFusion-L, BEV-MAE achieves new state-of-the-art LiDAR-based 3D object detection results, with 73.6 NDS and 69.6 mAP on the nuScenes benchmark. The source code will be released at https://github.com/VDIGPKU/BEV-MAE.

Real-time semantic segmentation via mutual optimization of spatial details and semantic information

Semantic segmentation is a pivotal task in the field of computer vision, encompassing diverse applications and undergoing continuous development. Despite the growing dominance of deep learning methods in this field, many existing network models suffer from trade-offs between accuracy and computational cost, or between speed and accuracy. In essence, semantic segmentation aims to extract semantic information from deep features and optimize them before upsampling output. However, shallow features tend to contain more detailed information but also more noise, while deep features have strong semantic information but lose some spatial information. To address this issue, we propose a novel mutual optimization strategy based on shallow spatial information and deep semantic information, and construct a details and semantic mutual optimization network (DSMONet). This effectively reduces the noise in the shallow features and guides the deep features to reconstruct the lost spatial information, avoiding cumbersome side auxiliary or complex decoders. The Mutual Optimization Module (MOM) includes Semantic Adjustment Details Module (SADM) and Detail Guided Semantic Module (DGSM), which enables mutual optimization of shallow spatial information and deep semantic information. Comparative evaluations against other methods demonstrate that DSMONet achieves a favorable balance between accuracy and speed. On the Cityscapes dataset, DSMONet achieves performances of 79.3% mean of class-wise intersection-over-union (mIoU)/44.6 frames per second (FPS) and 78.0% mIoU/102 FPS. The code is available at https://github.com/m828/DSMONet.

An approach for handwritten Chinese text recognition unifying character segmentation and recognition

Text line recognition methods can be categorized into explicit segmentation based and implicit segmentation based ones. Explicit segmentation based methods require character-level annotation during training, while implicit segmentation based methods, trained on line-level annotated data, face alignment drift challenges. Though some methods have been proposed to address these challenges using weakly supervised object detection, they often rely on cumbersome pseudo-box generation processes and complex decoding. In this paper, we propose a unified framework to overcome these challenges, achieving high accuracy in text recognition and character segmentation. To eliminate the need of character-level annotated real text line data in training, we introduce a novel training paradigm that utilizes character-level annotated synthetic data and line-level annotated real data jointly. For synthetic data, candidate characters are explicitly aligned with labeled characters to generate hard labels for supervising model training. For real data, implicit alignment is produced by Connectionist Temporal Classification (CTC) mapping to provide soft labels for weakly-supervised model training. And for inference, we propose two decoding strategies leveraging the advantages of Non-Maximum Suppression (NMS) and CTC decoding. Extensive experiments on benchmark datasets demonstrate the superior performance of our method in text recognition and character localization, even with minimal amounts of character-level annotated line data.

CRMEFNet: A coupled refinement, multiscale exploration and fusion network for medical image segmentation

Accurate segmentation of target areas in medical images, such as lesions, is essential for disease diagnosis and clinical analysis. In recent years, deep learning methods have been intensively researched and have generated significant progress in medical image segmentation tasks. However, most of the existing methods have limitations in modeling multilevel feature representations and identification of complex textured pixels at contrasting boundaries. This paper proposes a novel coupled refinement and multiscale exploration and fusion network (CRMEFNet) for medical image segmentation, which explores in the optimization and fusion of multiscale features to address the abovementioned limitations. The CRMEFNet consists of three main innovations: a coupled refinement module (CRM), a multiscale exploration and fusion module (MEFM), and a cascaded progressive decoder (CPD). The CRM decouples features into low-frequency body features and high-frequency edge features, and performs targeted optimization of both to enhance intraclass uniformity and interclass differentiation of features. The MEFM performs a two-stage exploration and fusion of multiscale features using our proposed multiscale aggregation attention mechanism, which explores the differentiated information within the cross-level features, and enhances the contextual connections between the features, to achieves adaptive feature fusion. Compared to existing complex decoders, the CPD decoder (consisting of the CRM and MEFM) can perform fine-grained pixel recognition while retaining complete semantic location information. It also has a simple design and excellent performance. The experimental results from five medical image segmentation tasks, ten datasets and twelve comparison models demonstrate the state-of-the-art performance, interpretability, flexibility and versatility of our CRMEFNet.

Rationalizing Graph Neural Networks with Data Augmentation

Graph rationales are representative subgraph structures that best explain and support the graph neural network (GNN) predictions. Graph rationalization involves the joint identification of these subgraphs during GNN training, resulting in improved interpretability and generalization. GNN is widely used for node-level tasks such as paper classification and graph-level tasks such as molecular property prediction. However, on both levels, little attention has been given to GNN rationalization and the lack of training examples makes it difficult to identify the optimal graph rationales. In this work, we address the problem by proposing a unified data augmentation framework with two novel operations on environment subgraphs to rationalize GNN prediction. We define the environment subgraph as the remaining subgraph after rationale identification and separation. The framework efficiently performs rationale–environment separation in the representation space for a node’s neighborhood graph or a graph’s complete structure to avoid the high complexity of explicit graph decoding and encoding. We conduct experiments on 17 datasets spanning node classification, graph classification, and graph regression. Results demonstrate that our framework is effective and efficient in rationalizing and enhancing GNNs for different levels of tasks on graphs.

Derived dynamic scheduling for belief propagation decoding of LDPC codes

AbstractAmong numerous dynamic scheduling strategies for low‐density parity‐check codes, many of them are of high complexity due to repetitive computation and ordering of belief residuals. To address these challenges, two novel approaches are proposed: the derived informed variable‐to‐check residual belief propagation and the derived variable‐node and variable‐to‐check‐edge residual belief propagation for low‐density parity‐check codes. Simulation results show that the derived dynamic scheduling exhibits a substantial gain of approximately 0.15 dB in signal‐to‐noise ratio over the non‐derived one with considerably lower decoding complexity when targeting a bit‐error rate of 10−4.