Author name cluster

Hanwen Liu

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

14 papers

2 author rows

AAAI Conference 2026 Conference Paper

HardF-SNN: Hardware-Friendly Quantization for Spiking Neural Networks with Efficient Integer-Arithmetic-Only Inference

Hanwen Liu
Kexin Shi
Jieyuan Zhang
Yimeng Shan
Jibin Wu
Wenyu Chen
Malu Zhang

Spiking Neural Networks (SNNs) are emerging as a promising energy-efficient alternative to Artificial Neural Networks (ANNs) due to their event-driven computation paradigm. However, recent advances toward large-scale high-performance SNNs inevitably lead to substantial memory and computational overhead. While quantization offers a potential way, many quantization approaches fail to deliver verifiable efficiency gains on resource-constrained hardware platforms. In this paper, we propose a lightweight and hardware-friendly SNN, termed HardF-SNN. Specifically, we first build a baseline model using shared-scale quantization and BN folding to simulate integer-only inference, as this has not been thoroughly discussed in prior SNN works. Then, through empirical and theoretical analysis, we identify that the baseline suffers from accuracy degradation and may cause training failure. To mitigate these issues, we propose proportional shared-scale quantization for enhanced dynamic range and integer-only BN using bit-shifting to stabilize training. Extensive experiments show that HardF-SNN achieves an optimal balance between performance and efficiency with excellent hardware compatibility. To demonstrate its effectiveness on resource-limited platforms, HardF-SNN is deployed on a dedicated FPGA-based hardware accelerator. Evaluation results indicate that our implementation achieves significant performance improvements over several existing hardware accelerators.

PDF Details DOI

AAAI Conference 2026 Conference Paper

Training-Free ANN-to-SNN Conversion for High-Performance Spiking Transformers

Jingya Wang
Xin Deng
Wenjie Wei
Dehao Zhang
Shuai Wang
Qian Sun
Jieyuan Zhang
Hanwen Liu

Leveraging the event-driven paradigm, Spiking Neural Networks (SNNs) offer a promising approach for constructing energy-efficient Transformer architectures. Compared to directly trained Spiking Transformers, ANN-to-SNN conversion methods bypass the high training costs. However, existing methods still suffer from notable limitations, failing to effectively handle nonlinear operations in Transformer architectures and requiring additional fine-tuning processes for pre-trained ANNs. To address these issues, we propose a high-performance and training-free ANN-to-SNN conversion framework tailored for Transformer architectures. Specifically, we introduce a Multi-basis Exponential Decay (MBE) neuron, which employs an exponential decay strategy and multi-basis encoding method to efficiently approximate various nonlinear operations. It removes the requirement for weight modifications in pre-trained ANNs. Extensive experiments across diverse tasks (CV, NLU, NLG) and mainstream Transformer architectures (ViT, RoBERTa, GPT-2) demonstrate that our method achieves near-lossless conversion accuracy with significantly lower latency. This provides a promising pathway for the efficient and scalable deployment of Spiking Transformers in real-world applications.

PDF Details DOI

NeurIPS Conference 2025 Conference Paper

Removing Concepts from Text-to-Image Models with Only Negative Samples

Hanwen Liu
Yadong Mu

This work introduces Clipout, a method for removing a target concept in pre-trained text-to-image models. By randomly clipping units from the learned data embedding and using a contrastive objective, models are encouraged to differentiate these clipped embedding vectors. Our goal is to remove private, copyrighted, inaccurate, or harmful concepts from trained models without the need for retraining. This is achieved by considering only negative samples and generating them in a bootstrapping-like manner, requiring minimal prior knowledge. Additionally, theoretical analyses are provided to further understand our proposed Clipout. Extensive experiments on text-to-image show that Clipout is simple yet highly effective and efficient compared with previous state-of-the-art approaches.

PDF Details

NeurIPS Conference 2025 Conference Paper

S$^2$NN: Sub-bit Spiking Neural Networks

Wenjie Wei
Malu Zhang
Jieyuan (Eric) Zhang
Ammar Belatreche
Shuai Wang
Yimeng Shan
Hanwen Liu
Honglin Cao

Spiking Neural Networks (SNNs) offer an energy-efficient paradigm for machine intelligence, but their continued scaling poses challenges for resource-limited deployment. Despite recent advances in binary SNNs, the storage and computational demands remain substantial for large-scale networks. To further explore the compression and acceleration potential of SNNs, we propose Sub-bit Spiking Neural Networks (S$^2$NNs) that represent weights with less than one bit. Specifically, we first establish an S$^2$NN baseline by leveraging the clustering patterns of kernels in well-trained binary SNNs. This baseline is highly efficient but suffers from \textit{outlier-induced codeword selection bias} during training. To mitigate this issue, we propose an \textit{outlier-aware sub-bit weight quantization} (OS-Quant) method, which optimizes codeword selection by identifying and adaptively scaling outliers. Furthermore, we propose a \textit{membrane potential-based feature distillation} (MPFD) method, improving the performance of highly compressed S$^2$NN via more precise guidance from a teacher model. Extensive results on vision reveal that S$^2$NN outperforms existing quantized SNNs in both performance and efficiency, making it promising for edge computing applications.

PDF Details

NeurIPS Conference 2025 Conference Paper

SALoM: Structure Aware Temporal Graph Networks with Long-Short Memory Updater

Hanwen Liu
Longjiao Zhang
Rui Wang
Tongya Zheng
Sai Wu
Chang Yao
Mingli Song

Dynamic graph learning is crucial for accurately modeling complex systems by integrating topological structure and temporal information within graphs. While memory-based methods are commonly used and excel at capturing short-range temporal correlations, they struggle with modeling long-range dependencies, harmonizing long-range and short-range correlations, and integrating structural information effectively. To address these challenges, we present SALoM: Structure Aware Temporal Graph Networks with Long-Short Memory Updater. SALoM features a memory module that addresses gradient vanishing and information forgetting, enabling the capture of long-term dependencies across various time scales. Additionally, SALoM utilizes a long-short memory updater (LSMU) to dynamically balance long-range and short-range temporal correlations, preventing over-generalization. By integrating co-occurrence encoding and LSMU through information bottleneck-based fusion, SALoM effectively captures both the structural and temporal information within graphs. Experimental results across various graph datasets demonstrate SALoM's superior performance, achieving state-of-the-art results in dynamic graph link prediction. Our code is openly accessible at https: //github. com/wave5418/SALoM.

PDF Details

NeurIPS Conference 2025 Conference Paper

Unveiling the Spatial-temporal Effective Receptive Fields of Spiking Neural Networks

Jieyuan (Eric) Zhang
Xiaolong Zhou
Shuai Wang
Wenjie Wei
Hanwen Liu
Qian Sun
Malu Zhang
Yang Yang

Spiking Neural Networks (SNNs) demonstrate significant potential for energy-efficient neuromorphic computing through an event-driven paradigm. While training methods and computational models have greatly advanced, SNNs struggle to achieve competitive performance in visual long-sequence modeling tasks. In artificial neural networks, the effective receptive field (ERF) serves as a valuable tool for analyzing feature extraction capabilities in visual long-sequence modeling. Inspired by this, we introduce the Spatio-Temporal Effective Receptive Field (ST-ERF) to analyze the ERF distributions across various Transformer-based SNNs. Based on the proposed ST-ERF, we reveal that these models suffer from establishing a robust global ST-ERF, thereby limiting their visual feature modeling capabilities. To overcome this issue, we propose two novel channel-mixer architectures: \underline{m}ulti-\underline{l}ayer-\underline{p}erceptron-based m\underline{ixer} (MLPixer) and \underline{s}plash-and-\underline{r}econstruct \underline{b}lock (SRB). These architectures enhance global spatial ERF through all timesteps in early network stages of Transformer-based SNNs, improving performance on challenging visual long-sequence modeling tasks. Extensive experiments conducted on the Meta-SDT variants and across object detection and semantic segmentation tasks further validate the effectiveness of our proposed method. Beyond these specific applications, we believe the proposed ST-ERF framework can provide valuable insights for designing and optimizing SNN architectures across a broader range of tasks. The code is available at \href{https: //github. com/EricZhang1412/Spatial-temporal-ERF}{\faGithub~EricZhang1412/Spatial-temporal-ERF}.

PDF Details

ECAI Conference 2024 Conference Paper

Reasonable Gradients for Online Training Algorithms in Spiking Neural Networks

Lang Xue
Hanwen Liu
Jing Wang
Hong Qu

Spiking neural networks (SNNs) have the potential to simulate sparse and spatio-temporal dynamics observed in biological neurons, making them promising for achieving energy-efficient artificial general intelligence. While backpropagation through time (BPTT) ensures reliable precision for training SNNs, it is hampered by high computation and storage complexity and does not conform to the instantaneous learning mechanism in brains. On the contrary, online training algorithms, which are biologically interpretable, offer low latency and memory efficiency, and are well-suited for on-chip learning applications. However, recent research exhibit a deficiency in the scientific comprehension of online gradients, which leads to certain limitations. To address this issue, we conduct an in-depth analysis of the calculation deviation in chain derivations induced by weight update and find two pivotal factors that affect the accuracy of online gradients: completeness and timeliness. To further enhance the performance of online training leveraging these findings, we propose spatio-temporal online learning (STOL), which substantially ameliorates the accuracy of the online gradients and demonstrates superior computation and memory efficiency. Our experiments on CIFAR-10, CIFAR-100, ImageNet, CIFAR10-DVS, and DVS128-Gesture datasets demonstrate that our method achieves state-of-the-art performance across most of these tasks. Besides, it shows a great improvement compared with existing online training algorithms.

Details

YNIMG Journal 2024 Journal Article

Ultra-high-resolution mapping of myelin and g-ratio in a panel of Mbp enhancer-edited mouse strains using microstructural MRI

Vladimir Grouza
Hooman Bagheri
Hanwen Liu
Marius Tuznik
Zhe Wu
Nicole Robinson
Katherine A. Siminovitch
Alan C. Peterson

Non-invasive myelin water fraction (MWF) and g-ratio mapping using microstructural MRI have the potential to offer critical insights into brain microstructure and our understanding of neuroplasticity and neuroinflammation. By leveraging a unique panel of variably hypomyelinating mouse strains, we validated a high-resolution, model-free image reconstruction method for whole-brain MWF mapping. Further, by employing a bipolar gradient echo MRI sequence, we achieved high spatial resolution and robust mapping of MWF and g-ratio across the whole mouse brain. Our regional white matter-tract specific analyses demonstrated a graded decrease in MWF in white matter tracts which correlated strongly with myelin basic protein gene (Mbp) mRNA levels. Using these measures, we derived the first sensitive calibrations between MWF and Mbp mRNA in the mouse. Minimal changes in axonal density supported our hypothesis that observed MWF alterations stem from hypomyelination. Overall, our work strongly emphasizes the potential of non-invasive, MRI-derived MWF and g-ratio modeling for both preclinical model validation and ultimately translation to humans.

Details DOI

ICML Conference 2023 Conference Paper

Trapdoor Normalization with Irreversible Ownership Verification

Hanwen Liu
Zhenyu Weng
Yuesheng Zhu
Yadong Mu

This paper introduces a deep model watermark with an irreversible ownership verification scheme: Trapdoor Normalization (TdN), inspired by the trapdoor function in traditional cryptography. To protect intellectual property within deep models, the proposed method is able to embed ownership information into normalization layers during training. We argue and empirically validate that relevant methods are vulnerable to ambiguity attacks, where the forged watermarks can cast ambiguity over the ownership verification. The primary trait that distinguishes this work from previous ones, is its design of a bidirectional connection between watermarks and deep models. Thereby, TdN enables an irreversible ownership verification scheme that is difficult for the adversary to compromise. In this way, the proposed TdN can effectively defeat ambiguity attacks. Extensive experiments demonstrate that the proposed method is not only superior to previous state-of-the-art methods in robustness, but also has better efficiency.

Details

YNIMG Journal 2022 Journal Article

Self-labelled encoder-decoder (SLED) for multi-echo gradient echo-based myelin water imaging

Hanwen Liu
Vladimir Grouza
Marius Tuznik
Katherine A. Siminovitch
Hooman Bagheri
Alan Peterson
David A. Rudko

PURPOSE: Reconstruction of high quality myelin water imaging (MWI) maps is challenging, particularly for data acquired using multi-echo gradient echo (mGRE) sequences. A non-linear least squares fitting (NLLS) approach has often been applied for MWI. However, this approach may produce maps with limited detail and, in some cases, sub-optimal signal to noise ratio (SNR), due to the nature of the voxel-wise fitting. In this study, we developed a novel, unsupervised learning method called self-labelled encoder-decoder (SLED) to improve gradient echo-based MWI data fitting. METHODS: Ultra-high resolution, MWI data was collected from five mouse brains with variable levels of myelination, using a mGRE sequence. Imaging data was acquired using a 7T preclinical MRI system. A self-labelled, encoder-decoder network was implemented in TensorFlow for calculation of myelin water fraction (MWF) based on the mGRE signal decay. A simulated MWI phantom was also created to evaluate the performance of MWF estimation. RESULTS: Compared to NLLS, SLED demonstrated improved MWF estimation, in terms of both stability and accuracy in phantom tests. In addition, SLED produced less noisy MWF maps from high resolution MR microscopy images of mouse brain tissue. It specifically resulted in lower noise amplification for all mouse genotypes that were imaged and yielded mean MWF values in white matter ROIs that were highly correlated with those derived from standard NLLS fitting. Lastly, SLED also exhibited higher tolerance to low SNR data. CONCLUSION: Due to its unsupervised and self-labeling nature, SLED offers a unique alternative to analyze gradient echo-based MWI data, providing accurate and stable MWF estimations.

Details DOI

ICML Conference 2021 Conference Paper

Watermarking Deep Neural Networks with Greedy Residuals

Hanwen Liu
Zhenyu Weng
Yuesheng Zhu

Deep neural networks (DNNs) are considered as intellectual property of their corresponding owners and thus are in urgent need of ownership protection, due to the massive amount of time and resources invested in designing, tuning and training them. In this paper, we propose a novel watermark-based ownership protection method by using the residuals of important parameters. Different from other watermark-based ownership protection methods that rely on some specific neural network architectures and during verification require external data source, namely ownership indicators, our method does not explicitly use ownership indicators for verification to defeat various attacks against DNN watermarks. Specifically, we greedily select a few and important model parameters for embedding so that the impairment caused by the changed parameters can be reduced and the robustness against different attacks can be improved as the selected parameters can well preserve the model information. Also, without the external data sources for verification, the adversary can hardly cast doubts on ownership verification by forging counterfeit watermarks. The extensive experiments show that our method outperforms previous state-of-the-art methods in five tasks.

Details

YNIMG Journal 2020 Journal Article

Myelin water imaging data analysis in less than one minute

Hanwen Liu
Qing-San Xiang
Roger Tam
Adam V. Dvorak
Alex L. MacKay
Shannon H. Kolind
Anthony Traboulsee
Irene M. Vavasour

Purpose Based on a deep learning neural network (NN) algorithm, a super fast and easy to implement data analysis method was proposed for myelin water imaging (MWI) to calculate the myelin water fraction (MWF). Methods A NN was constructed and trained on MWI data acquired by a 32-echo 3D gradient and spin echo (GRASE) sequence. Ground truth labels were created by regularized non-negative least squares (NNLS) with stimulated echo corrections. Voxel-wise GRASE data from 5 brains (4 healthy, 1 multiple sclerosis (MS)) were used for NN training. The trained NN was tested on 2 healthy brains, 1 MS brain with segmented lesions, 1 healthy spinal cord, and 1 healthy brain acquired from a different scanner. Results Production of whole brain MWF maps in approximately 33 s can be achieved by a trained NN without graphics card acceleration. For all testing regions, no visual differences between NN and NNLS MWF maps were observed, and no obvious regional biases were found. Quantitatively, all voxels exhibited excellent agreement between NN and NNLS (all R2>0. 98, p < 0. 001, mean absolute error <0. 01). Conclusion The time for accurate MWF calculation can be dramatically reduced to less than 1 min by the proposed NN, addressing one of the barriers facing future clinical feasibility of MWI.

Details DOI

AAAI Conference 2019 Conference Paper

Multigrid Backprojection Super–Resolution and Deep Filter Visualization

Pablo Navarrete Michelini
Hanwen Liu
Dan Zhu

We introduce a novel deep–learning architecture for image upscaling by large factors (e. g. 4×, 8×) based on examples of pristine high–resolution images. Our target is to reconstruct high–resolution images from their downscale versions. The proposed system performs a multi–level progressive upscaling, starting from small factors (2×) and updating for higher factors (4× and 8×). The system is recursive as it repeats the same procedure at each level. It is also residual since we use the network to update the outputs of a classic upscaler. The network residuals are improved by Iterative Back–Projections (IBP) computed in the features of a convolutional network. To work in multiple levels we extend the standard back– projection algorithm using a recursion analogous to Multi– Grid algorithms commonly used as solvers of large systems of linear equations. We finally show how the network can be interpreted as a standard upsampling–and–filter upscaler with a space–variant filter that adapts to the geometry. This approach allows us to visualize how the network learns to upscale. Finally, our system reaches state of the art quality for models with relatively few number of parameters.

PDF Details

YNICL Journal 2019 Journal Article

Rapid myelin water imaging for the assessment of cervical spinal cord myelin damage

Adam V. Dvorak
Emil Ljungberg
Irene M. Vavasour
Hanwen Liu
Poljanka Johnson
Alexander Rauscher
John L.K. Kramer
Roger Tam

BACKGROUND: ). OBJECTIVE: metrics. METHODS: data were acquired in 2 PLS, 1 relapsing-remitting MS (RRMS), 1 primary-progressive MS (PPMS) and 2 NMO subjects, as well as 6 age (±3 yrs) and sex matched healthy controls (HC). Internal cord structure guided template registrations, used for region of interest (ROI) analysis. Z score maps were calculated for the difference between disease subject and mean HC metric values. RESULTS: metrics showed similar trends, corroborating the MWF results and providing complementary information. CONCLUSION: measures.

Details DOI