Sheng Chang

Tai An, Ruwu Cai, Yanzhe Zhang et al.

In the era of large language models (LLMs), N:M sparsity has emerged as a structured compression technique critical for accelerating inference. While prior work has primarily focused on weight sparsity, it often suffers from significant accuracy degradation. Activation sparsity, though promising, is typically training-dependent and faces challenges in generalization. To address these limitations, we introduce Amber Pruner, a training-free N:M activation sparsity method designed specifically for the prefill stage, targeting the acceleration of linear projection layers in LLMs. Extensive experiments across multiple models and sparsity ratios (2:4, 4:8, and 8:16) demonstrate that Amber Pruner can effectively sparsify and accelerate more than 55% of linear computations without requiring model retraining. To further enhance generality and efficiency, we propose Outstanding-sparse, a unified framework that integrates Amber Pruner with post-training W8A8 quantization. Our approach preserves strong performance across a range of downstream tasks, with notable advantages in generative tasks. This work pioneers a new frontier in activation sparsity, providing foundational insights that are poised to guide the co-evolution of algorithms and architectures in the design of next-generation AI systems.

Zhifei Shi, Zongyao Yin, Sheng Chang et al.

Achieving a balance between computational efficiency and detection accuracy in the realm of rotated bounding box object detection within aerial imagery is a significant challenge. While prior research has aimed at creating lightweight models that enhance computational performance and feature extraction, there remains a gap in the performance of these networks when it comes to the detection of small and multi-scale objects in remote sensing (RS) imagery. To address these challenges, we present a novel enhancement to the YOLOv8 model, tailored for oriented object detection tasks and optimized for environments with limited computational resources. Our model features a wavelet transform-based C2f module for capturing associative features and an Adaptive Scale Feature Pyramid (ASFP) module that leverages P2 layer details. Additionally, the incorporation of GhostDynamicConv significantly contributes to the model's lightweight nature, ensuring high efficiency in aerial imagery analysis. Featuring a parameter count of 21.6M, our approach provides a more efficient architectural design than DecoupleNet, which has 23.3M parameters, all while maintaining detection accuracy. On the DOTAv1.0 dataset, our model demonstrates a mean Average Precision (mAP) that is competitive with leading methods such as DecoupleNet. The model's efficiency, combined with its reduced parameter count, makes it a strong candidate for aerial object detection, particularly in resource-constrained environments.

Huaicheng Zhang, Ruoxin Wang, Chenlian Zhou et al.

As one of the most effective methods for cardiovascular disease (CVD) diagnosis, multi-lead Electrocardiogram (ECG) signals present a characteristic multi-sensor information fusion challenge that has been continuously researched in deep learning domains. Despite the numerous algorithms proposed with different DL architectures, maintaining a balance among performance, computational complexity, and multi-source ECG feature fusion remains challenging. Recently, state space models (SSMs), particularly Mamba, have demonstrated remarkable effectiveness across various fields. Their inherent design for high-efficiency computation and linear complexity makes them particularly suitable for low-dimensional data like ECGs. This work proposes S2M2ECG, an SSM architecture featuring three-level fusion mechanisms: (1) Spatio-temporal bi-directional SSMs with segment tokenization for low-level signal fusion, (2) Intra-lead temporal information fusion with bi-directional scanning to enhance recognition accuracy in both forward and backward directions, (3) Cross-lead feature interaction modules for spatial information fusion. To fully leverage the ECG-specific multi-lead mechanisms inherent in ECG signals, a multi-branch design and lead fusion modules are incorporated, enabling individual analysis of each lead while ensuring seamless integration with others. Experimental results reveal that S2M2ECG achieves superior performance in the rhythmic, morphological, and clinical scenarios. Moreover, its lightweight architecture ensures it has nearly the fewest parameters among existing models, making it highly suitable for efficient inference and convenient deployment. Collectively, S2M2ECG offers a promising alternative that strikes an excellent balance among performance, computational complexity, and ECG-specific characteristics, paving the way for high-performance, lightweight computations in CVD diagnosis.

Yue Shi, Liangxiu Han, Lianghao Han et al.

Realistic hyperspectral image (HSI) super-resolution (SR) techniques aim to generate a high-resolution (HR) HSI with higher spectral and spatial fidelity from its low-resolution (LR) counterpart. The generative adversarial network (GAN) has proven to be an effective deep learning framework for image super-resolution. However, the optimisation process of existing GAN-based models frequently suffers from the problem of mode collapse, leading to the limited capacity of spectral-spatial invariant reconstruction. This may cause the spectral-spatial distortion on the generated HSI, especially with a large upscaling factor. To alleviate the problem of mode collapse, this work has proposed a novel GAN model coupled with a latent encoder (LE-GAN), which can map the generated spectral-spatial features from the image space to the latent space and produce a coupling component to regularise the generated samples. Essentially, we treat an HSI as a high-dimensional manifold embedded in a latent space. Thus, the optimisation of GAN models is converted to the problem of learning the distributions of high-resolution HSI samples in the latent space, making the distributions of the generated super-resolution HSIs closer to those of their original high-resolution counterparts. We have conducted experimental evaluations on the model performance of super-resolution and its capability in alleviating mode collapse. The proposed approach has been tested and validated based on two real HSI datasets with different sensors (i.e. AVIRIS and UHD-185) for various upscaling factors and added noise levels, and compared with the state-of-the-art super-resolution models (i.e. HyCoNet, LTTR, BAGAN, SR- GAN, WGAN).

Yue Shi, Liangxiu Han, Anthony Kleerekoper et al.

Late blight disease is one of the most destructive diseases in potato crop, leading to serious yield losses globally. Accurate diagnosis of the disease at early stage is critical for precision disease control and management. Current farm practices in crop disease diagnosis are based on manual visual inspection, which is costly, time consuming, subject to individual bias. Recent advances in imaging sensors (e.g. RGB, multiple spectral and hyperspectral cameras), remote sensing and machine learning offer the opportunity to address this challenge. Particularly, hyperspectral imagery (HSI) combining with machine learning/deep learning approaches is preferable for accurately identifying specific plant diseases because the HSI consists of a wide range of high-quality reflectance information beyond human vision, capable of capturing both spectral-spatial information. The proposed method considers the potential disease specific reflectance radiation variance caused by the canopy structural diversity, introduces the multiple capsule layers to model the hierarchical structure of the spectral-spatial disease attributes with the encapsulated features to represent the various classes and the rotation invariance of the disease attributes in the feature space. We have evaluated the proposed method with the real UAV-based HSI data under the controlled field conditions. The effectiveness of the hierarchical features has been quantitatively assessed and compared with the existing representative machine learning/deep learning methods. The experiment results show that the proposed model significantly improves the accuracy performance when considering hierarchical-structure of spectral-spatial features, comparing to the existing methods only using spectral, or spatial or spectral-spatial features without consider hierarchical-structure of spectral-spatial features.

Yue Shi, Liangxiu Han, Wenjiang Huang et al.

Spectral-spatial based deep learning models have recently proven to be effective in hyperspectral image (HSI) classification for various earth monitoring applications such as land cover classification and agricultural monitoring. However, due to the nature of "black-box" model representation, how to explain and interpret the learning process and the model decision, especially for vegetation classification, remains an open challenge. This study proposes a novel interpretable deep learning model -- a biologically interpretable two-stage deep neural network (BIT-DNN), by incorporating the prior-knowledge (i.e. biophysical and biochemical attributes and their hierarchical structures of target entities) based spectral-spatial feature transformation into the proposed framework, capable of achieving both high accuracy and interpretability on HSI based classification tasks. The proposed model introduces a two-stage feature learning process: in the first stage, an enhanced interpretable feature block extracts the low-level spectral features associated with the biophysical and biochemical attributes of target entities; and in the second stage, an interpretable capsule block extracts and encapsulates the high-level joint spectral-spatial features representing the hierarchical structure of biophysical and biochemical attributes of these target entities, which provides the model an improved performance on classification and intrinsic interpretability with reduced computational complexity. We have tested and evaluated the model using four real HSI datasets for four separate tasks (i.e. plant species classification, land cover classification, urban scene recognition, and crop disease recognition tasks). The proposed model has been compared with five state-of-the-art deep learning models.

Ruihan Hu, Qijun Huang, Sheng Chang et al.

Machine learning algorithms have been effectively applied into various real world tasks. However, it is difficult to provide high-quality machine learning solutions to accommodate an unknown distribution of input datasets; this difficulty is called the uncertainty prediction problems. In this paper, a margin-based Pareto deep ensemble pruning (MBPEP) model is proposed. It achieves the high-quality uncertainty estimation with a small value of the prediction interval width (MPIW) and a high confidence of prediction interval coverage probability (PICP) by using deep ensemble networks. In addition to these networks, unique loss functions are proposed, and these functions make the sub-learners available for standard gradient descent learning. Furthermore, the margin criterion fine-tuning-based Pareto pruning method is introduced to optimize the ensembles. Several experiments including predicting uncertainties of classification and regression are conducted to analyze the performance of MBPEP. The experimental results show that MBPEP achieves a small interval width and a low learning error with an optimal number of ensembles. For the real-world problems, MBPEP performs well on input datasets with unknown distributions datasets incomings and improves learning performance on a multi task problem when compared to that of each single model.

Zhiri Tang, Ruohua Zhu, Peng Lin et al.

Memristive neural networks (MNNs), which use memristors as neurons or synapses, have become a hot research topic recently. However, most memristors are not compatible with mainstream integrated circuit technology and their stabilities in large-scale are not very well so far. In this paper, a hardware friendly MNN circuit is introduced, in which the memristive characteristics are implemented by digital integrated circuit. Through this method, spike timing dependent plasticity (STDP) and unsupervised learning are realized. A weight sharing mechanism is proposed to bridge the gap of network scale and hardware resource. Experiment results show the hardware resource is significantly saved with it, maintaining good recognition accuracy and high speed. Moreover, the tendency of resource increase is slower than the expansion of network scale, which infers our method's potential on large scale neuromorphic network's realization.