arXiv subjects

Yujiang Wang

Publications and source records attributed to Yujiang Wang.

At least 19 recordsLinked to original sources

arXiv2026

Habitual lifestyle timing explains circadian timing, but daily lifestyle changes do not, in free-living humans across 2000 days

Background: Both between- and within-subject variations in circadian timing matter for health. If lifestyle changes could be used to regulate circadian timing, they would offer accessible and scalable routes to chronotherapy, but this link remains unclear under real-life conditions. Here, we explore how lifestyle 'traits' (such as typical wake time) and 'states' (day-to-day deviations from traits, such as waking up later than typical) explain between- and within-subject variation in acrophase (peak time) of the circadian rhythm of heart rate (CRHR). Methods: We collected free-living wearable data (smartwatch, continuous glucose monitor) from healthy volunteers for up to 4 weeks. The CRHR was derived from activity-adjusted heart rate, and acrophase was defined as time-of-day at daily CRHR peak. Sleep, food, and physical activity 'factors' were calculated and split into traits and states. Using a linear mixed-effects model, we tested how traits and states associate with between- and within-subject acrophase variance. Findings: Data from 105 healthy volunteers (66 female, age = 42.5 $\pm$ 15.7 years) spanning ~2000 days (18.8 $\pm$ 8.30 days each) were analysed. Traits were substantially more influential than states, explaining 42.3% versus 0.9% of total acrophase variance. Accordingly, traits explained 86.5% of between-subject variance, whereas states explained only 1.8% of within-subject variance. Sleep, food and physical activity factors contributed both jointly and uniquely, and lifestyle timing mattered most. Interpretation: Between-subject lifestyle traits explained acrophase better than within-subject lifestyle states. This asymmetry, alongside the considerable overlap between factors, supports sustained, holistic, timing-focused lifestyle adjustments as chronotherapy targets, testable through future interventional studies.

q-bio.QM↗

arXiv2026

Bayesian inference of sparsity in stable vector autoregressive processes

Advances in sensing technology have made it possible to collect large volumes of high-dimensional time-series data. In fields like genetics and neuroscience, key questions concern whether directed relationships between variables can be learned from these data. To this end, graphical vector autoregressions are a popular tool because zeros among the autoregressive coefficients and error precision matrix have natural interpretations in terms of Granger non-causality and contemporaneous conditional independence. In applications where system dynamics are subject to functional or structural constraints, assuming the process is stable can be advantageous. However, enforcing stability demands restricting the autoregressive coefficients to lie in a constrained space with a complex geometry called the stationary region. The resulting inferential challenges are compounded when sparsity is also a requirement. Working in the Bayesian paradigm, we tackle the problem of developing a prior that simultaneously enforces stationarity and sparsity through parameter expansion, constructing a spike-and-slab prior with support constrained to the stationary region. A mixture of G-Wishart distributions provides a sparse prior for the error precision matrix. Computational inference is carried out using Metropolis-within-Gibbs, exploiting the No-U-Turn Sampler and reversible-jump steps. We demonstrate the inferential and predictive benefits of our approach through simulations and applications in macroeconomics and neuroscience.

stat.ME↗

arXiv2026

The Curse and Blessing of Mean Bias in FP4-Quantized LLM Training

FP4 training promises substantial memory and compute savings for large language models, but remains fragile because blockwise quantization is dictated by extreme activation magnitudes, which inflate dynamic range and compress long-tail signals. We identify a counterintuitive source of this failure: dominant activation outliers are not merely arbitrary sparse events, but are largely induced by a coherent rank-one mean bias, whose direction aligns with the leading anisotropic spectral component. This mean component strengthens during training, is amplified and reshaped by attention and FFN operators, and increasingly dominates top activation magnitudes. Crucially, this discovery reveals that a seemingly complex outlier-suppression problem admits a truly simple solution: isolate the coherent mean before quantization. We therefore propose Averis, a mean-residual splitting quantization method that separates the mean component using only reductions and elementwise subtractions before FP4 quantization. Across Qwen3 0.6B Dense trained on 100B tokens and Qwen3 7B A1.5B MoE trained on 50B tokens, Averis enables robust W4A4G4 FP4 training, reducing BF16 loss gaps to 1.19%/0.81% versus 2.05%/1.10% for NVIDIA's recently released Hadamard-based outlier-smoothing method, while limiting downstream gaps to 0.89/0.71 points. With only 2.20% end-to-end overhead over vanilla NVFP4, about 30% of NVIDIA's Hadamard-based design, Averis provides a hardware-efficient path to stable low-bit LLM training. Complementary to Hadamard, Averis further reduces the Qwen3-0.6B loss and downstream gaps to 0.94% and 0.73 points when combined. Code is available at: https://anonymous.4open.science/r/averis-504D.

cs.LG↗

arXiv2026

UniG2U-Bench: Do Unified Models Advance Multimodal Understanding?

Unified multimodal models have recently demonstrated strong generative capabilities, yet whether and when generation improves understanding remains unclear. Existing benchmarks lack a systematic exploration of the specific tasks where generation facilitates understanding. To this end, we introduce UniG2U-Bench, a comprehensive benchmark categorizing generation-to-understanding (G2U) evaluation into 7 regimes and 30 subtasks, requiring varying degrees of implicit or explicit visual transformations. Extensive evaluation of over 30 models reveals three core findings: 1) Unified models generally underperform their base Vision-Language Models (VLMs), and Generate-then-Answer (GtA) inference typically degrades performance relative to direct inference. 2) Consistent enhancements emerge in spatial intelligence, visual illusions, or multi-round reasoning subtasks, where enhanced spatial and shape perception, as well as multi-step intermediate image states, prove beneficial. 3) Tasks with similar reasoning structures and models sharing architectures exhibit correlated behaviors, suggesting that generation-understanding coupling induces class-consistent inductive biases over tasks, pretraining data, and model architectures. These findings highlight the necessity for more diverse training data and novel paradigms to fully unlock the potential of unified multimodal modeling.

cs.CV↗

arXiv2026

SD-MoE: Spectral Decomposition for Effective Expert Specialization

Mixture-of-Experts (MoE) architectures scale Large Language Models via expert specialization induced by conditional computation. In practice, however, expert specialization often fails: some experts become functionally similar, while others functioning as de facto shared experts, limiting the effective capacity and model performance. In this work, we analysis from a spectral perspective on parameter and gradient spaces, uncover that (1) experts share highly overlapping dominant spectral components in their parameters, (2) dominant gradient subspaces are strongly aligned across experts, driven by ubiquitous low-rank structure in human corpus, and (3) gating mechanisms preferentially route inputs along these dominant directions, further limiting specialization. To address this, we propose Spectral-Decoupled MoE (SD-MoE), which decomposes both parameter and gradient in the spectral space. SD-MoE improves performance across downstream tasks, enables effective expert specialization, incurring minimal additional computation, and can be seamlessly integrated into a wide range of existing MoE architectures, including Qwen and DeepSeek.

cs.LG↗

arXiv2026

Multi-Head Attention as a Source of Catastrophic Forgetting in MoE Transformers

Mixture-of-Experts (MoE) architectures are often considered a natural fit for continual learning because sparse routing should localize updates and reduce interference, yet MoE Transformers still forget substantially even with sparse, well-balanced expert utilization. We attribute this gap to a pre-routing bottleneck: multi-head attention concatenates head-specific signals into a single post-attention router input, forcing routing to act on co-occurring feature compositions rather than separable head channels. We show that this router input simultaneously encodes multiple separately decodable semantic and structural factors with uneven head support, and that different feature compositions induce weakly aligned parameter-gradient directions; as a result, routing maps many distinct compositions to the same route. We quantify this collision effect via a route-wise effective composition number $N_{eff}$ and find that higher $N_{eff}$ is associated with larger old-task loss increases after continual training. Motivated by these findings, we propose MH-MoE, which performs head-wise routing over sub-representations to increase routing granularity and reduce composition collisions. On TRACE with Qwen3-0.6B/8B, MH-MoE effectively mitigates forgetting, reducing BWT on Qwen3-0.6B from 11.2% (LoRAMoE) to 4.5%.

cs.LG↗

arXiv2026

Open diffusion MRI and connectivity data for epilepsy and surgery: The IDEAS II release

Epileptic seizures are generated in cerebral networks that propagate ictal and interictal activity. The structure of cerebral networks underpinning epileptic activity can be inferred from diffusion-weighted MRI (DWI). However, publicly available DWI data in individuals with epilepsy are scarce, and processing is technically challenging due to scan-specific artifacts, limiting research progress. Here, we release raw DWI data from 216 individuals with epilepsy and 98 healthy controls. Subject identifiers align with our previous data release (IDEAS), which includes T1-weighted and FLAIR MRI, surgical details, and long-term seizure outcomes after surgery. Preprocessing reduced distortions and artifacts, while fully processed data include diffusion metric maps in native and template space. We also provide parcellated structural connectomes using multiple atlases and connectivity measures. To illustrate the utility of this IDEAS II data, we replicated ENIGMA consortium findings, observing widespread reductions of fractional anisotropy, particularly ipsilateral to the area of seizure onset. We further demonstrate localised abnormality, and network connectivity using streamline tractography in a patient who subsequently underwent temporal lobe resection. This open dataset offers a comprehensive resource to advance research on structural connectivity and surgical outcomes in epilepsy.

q-bio.NC↗

arXiv2026

Dispelling the Curse of Singularities in Neural Network Optimizations

This work investigates the optimization instability of deep neural networks from a less-explored yet insightful perspective: the emergence and amplification of singularities in the parametric space. Our analysis reveals that parametric singularities inevitably grow with gradient updates and further intensify alignment with representations, leading to increased singularities in the representation space. We show that the gradient Frobenius norms are bounded by the top singular values of the weight matrices, and as training progresses, the mutually reinforcing growth of weight and representation singularities, termed the curse of singularities, relaxes these bounds, escalating the risk of sharp loss explosions. To counter this, we propose Parametric Singularity Smoothing (PSS), a lightweight, flexible, and effective method for smoothing the singular spectra of weight matrices. Extensive experiments across diverse datasets, architectures, and optimizers demonstrate that PSS mitigates instability, restores trainability even after failure, and improves both training efficiency and generalization.

cs.LG↗

arXiv2026

Spectra: Rethinking Optimizers for LLMs Under Spectral Anisotropy

Gradient signals in LLM training are highly anisotropic: recurrent linguistic structure concentrates energy into a small set of dominant spectral directions, while context specific information resides in a long tail. We show that this spike tail separation persists throughout training, with the spike occupying only about 1.5% of directions yet dominating optimizer statistics. This dominance suppresses tail learning by contracting tail updates through second moment normalization and tightening the globally stable learning rate bound. Motivated by this analysis, we propose Spectra, a spike aware optimizer that suppresses the dominant low rank spike subspace without amplifying the noise sensitive spectral tail. Spectra tracks the spike subspace via cached, warm started power iteration and applies low rank spectral shaping with negligible overhead and substantially reduced optimizer state memory. On LLaMA3 8B trained on 50B tokens, Spectra reaches the same target loss 30% faster than AdamW, reduces per step end to end overhead by 0.7%, cuts optimizer state memory by 49.25%, and improves average downstream accuracy by 1.62%. Compared to Muon, Spectra is 5.1x faster in optimizer processing time, achieves a lower final loss, and improves average accuracy by 0.66%.

cs.LG↗

arXiv2025

Normative Modelling in Neuroimaging: A Practical Guide for Researchers

Normative modelling is an increasingly common statistical technique in neuroimaging that estimates population-level benchmarks in brain structure. It enables the quantification of individual deviations from expected distributions whilst accounting for biological and technical covariates without requiring large, matched control groups. This makes it a powerful alternative to traditional case-control studies for identifying brain structural alterations associated with pathology. Despite the availability of numerous modelling approaches and several toolboxes with pre-trained models, the distinct strengths and limitations of normative modelling make it difficult to determine how and when to implement them appropriately. This review offers practical guidance and outlines statistical considerations for clinical researchers using normative modelling in neuroimaging. Through a worked example using clinical epilepsy data, we outline considerations for responsible implementation of pre-trained normative models, to support their broad and rigorous adoption in neuroimaging research.

q-bio.NC↗

arXiv2025

Metis: Training LLMs with FP4 Quantization

This work identifies anisotropy in the singular value spectra of parameters, activations, and gradients as the fundamental barrier to low-bit training of large language models (LLMs). These spectra are dominated by a small fraction of large singular values, inducing wide numerical ranges that cause quantization bias and severe spectral distortion, ultimately degrading training performance. This work presents Metis, a spectral-domain quantization framework that partitions anisotropic spectra into narrower sub-distributions for independent quantization, thereby reducing errors and preserving spectral structure. To minimize overhead, Metis leverages two key properties of the dominant spectral subspace: preservation via sparsely random sampling and preservation via random projection, reducing decomposition cost to a negligible level. On LLaMA-3 8B trained with 100B tokens, Metis enables robust W4A4G4 training with FP4 quantization of weights, activations, and gradients, yielding only a 0.4% training loss gap and a 0.1% degradation in downstream accuracy relative to BF16. Beyond matching BF16 fidelity, Metis also surpasses our implementation of Nvidia's recently announced (yet to be publicly released) FP4 recipe, consistently achieving lower loss and higher downstream accuracy while incurring significantly lower computational overhead. The code implementation for Metis is available at: https://anonymous.4open.science/r/Metis-quantization-644B.

cs.LG↗

arXiv2025

Multimodal Modeling of Ultradian Rhythms Using the Hankel Alternative View of Koopman (HAVOK) Analysis

Ultradian rhythms - quasi-rhythmic fluctuations in behavior and physiology with periods shorter than 24 hours - are observed across various organisms, including humans. Despite their role in key biological processes such as sleep architecture and hormone regulation, their underlying mechanisms remain poorly understood. Here, we leveraged wearable sensor technology for continuous monitoring of physiological signals in 16 healthy participants over two weeks. By systematically removing circadian and longer-scale rhythms, we isolated ultradian dynamics and modeled them using the Hankel Alternative View of Koopman (HAVOK) framework,a data-driven approach based on Takens' embedding theorem and Koopman operator theory. This allowed us to characterize ultradian rhythms as an intermittently forced linear system and distinguish between regular oscillatory behavior and more complex dynamics. Across participants, ultradian fluctuations were well-described by the HAVOK model, with intermittent forcing consistently observed. The model demonstrated strong forecasting accuracy, with root mean squared error (RMSE) of $0.0315 \pm 0.02$, $0.0306 \pm 0.02$, and $0.0218 \pm 0.02$ in the leading time-delay coordinates. Notably, a significant sex difference in model rank (z = -2.06, p = 0.0396) suggests that sex hormones may play a key role in ultradian dynamics. These findings provide evidence for intermittently forced linear systems as a useful framework for understanding ultradian rhythms and their regulation.

q-bio.QM↗

arXiv2025

Seizure duration is associated with multiple timescales in interictal iEEG band power

Background Seizure severity can change from one seizure to the next within individual people with epilepsy. It is unclear if and how seizure severity is modulated over longer timescales. Characterising seizure severity variability over time could lead to tailored treatments. In this study, we test if continuously-recorded interictal intracranial EEG (iEEG) features encapsulate signatures of such modulations. Methods We analysed 20 subjects with iEEG recordings of at least one day. We identified cycles on timescales of hours to days embedded in long-term iEEG band power and associated them with seizure severity, which we approximated using seizure duration. In order to quantify these associations, we created linear-circular statistical models of seizure duration that incorporated different band power cycles within each subject. Findings In most subjects, seizure duration was weakly to moderately correlated with individual band power cycles. Combinations of multiple band power cycles significantly explained most of the variability in seizure duration. Specifically, we found 70% of the models had a higher than 60% adjusted $R^2$ across all subjects. From these models, around 80% were deemed to be above chance-level (p-value < 0.05) based on permutation tests. Models included cycles of ultradian, circadian and slower timescales in a subject-specific manner. Interpretation These results suggest that seizure severity, as measured by seizure duration, may be modulated over timescales of minutes to days by subject-specific cycles in interictal iEEG signal properties. These cycles likely serve as markers of seizure modulating processes. Future work can investigate biological drivers of these detected fluctuations and may inform novel treatment strategies that minimise seizure severity.

q-bio.NC↗

arXiv2025

Dual mechanism of Anti-Seizure Medications in controlling seizure activity

Background: Anti-seizure medications (ASMs) can reduce seizure duration, but their precise modes of action are unclear. Specifically, it is unknown whether ASMs shorten seizures by simply compressing existing seizure activity into a shorter time frame or by selectively suppressing certain seizure activity patterns. Methods: We analysed intracranial EEG (iEEG) recordings of 457 seizures from 28 people with epilepsy undergoing ASM tapering. Beyond measuring seizure occurrence and duration, we categorized distinct seizure activity patterns (states) based on spatial and frequency power characteristics and related these to different ASM levels. Results: We found that reducing ASM levels led to increased seizure frequency (r = 0.87, p < 0.001) and longer seizure duration ($\beta$ = -0.033, p < 0.001), consistent with prior research. Further analysis revealed two distinct mechanisms in which seizures became prolonged: Emergence of new seizure patterns - In approx. 40% of patients, ASM tapering unmasked additional seizure activity states, and seizures containing these 'taper-emergent states' were substantially longer (r = 0.49, p < 0.001). Prolongation of existing seizure patterns - Even in seizures without taper-emergent states, lower ASM levels still resulted in approx. 12-224% longer durations depending on the ASM dosage and tapering ($\beta$ = -0.049, p < 0.001). Conclusion: ASMs influence seizures through two mechanisms: they (i) suppress specific seizure activity patterns (states) in an all-or-nothing fashion and (ii) curtail the duration of other seizure patterns. These findings highlight the complex role of ASMs in seizure modulation and could inform personalized dosing strategies for epilepsy management. These findings may also have implications in understanding the effects of ASMs on cognition and mood.

q-bio.NC↗

arXiv2025

Combined impact of grey and superficial white matter abnormalities: implications for epilepsy surgery

Drug-resistant focal epilepsy is associated with abnormalities in the brain in both grey matter (GM) and superficial white matter (SWM). However, it is unknown if both types of abnormalities are important in supporting seizures. Here, we test if surgical removal of GM and/or SWM abnormalities relates to post-surgical seizure outcome in people with temporal lobe epilepsy (TLE). We analyzed structural imaging data from 143 TLE patients (pre-op dMRI and pre-op T1-weighted MRI) and 97 healthy controls. We calculated GM volume abnormalities and SWM mean diffusivity abnormalities and evaluated if their surgical removal distinguished seizure outcome groups post-surgically. At a group level, GM and SWM abnormalities were most common in the ipsilateral temporal lobe and hippocampus in people with TLE. Analyzing both modalities together, compared to in isolation, improved surgical outcome discrimination (GM AUC = 0.68, p < 0.01, WM AUC = 0.65, p < 0.01; Union AUC = 0.72, p < 0.01, Concordance AUC = 0.64, p = 0.04). Additionally, 100% of people who had all concordant abnormal regions resected had ILAE$_{1,2}$ outcomes. These findings suggest that regions identified as abnormal from both diffusion-weighted and T1-weighted MRIs are involved in the epileptogenic network and that resection of both types of abnormalities may enhance the chances of living without disabling seizures.

q-bio.NC↗

arXiv2025

Automated generation of epilepsy surgery resection masks; The RAMPS pipeline

MRI-based delineation of brain tissue removed by epilepsy surgery can be challenging due to post-operative brain shift. In consequence, most studies use manual approaches which are prohibitively time-consuming for large sample sizes, require expertise, and can be prone to errors. We propose RAMPS (Resections And Masks in Preoperative Space), an automated pipeline to generate a 3D resection mask of pre-operative tissue. Our pipeline leverages existing software including FreeSurfer, SynthStrip, Sythnseg and ANTS to generate a mask in the same space as the patient's pre-operative T1 weighted MRI. We compare our automated masks against manually drawn masks and two other existing pipelines (Epic-CHOP and ResectVol). Comparing to manual masks (N=87), RAMPS achieved a median(IQR) dice similarity of 0.86(0.078) in temporal lobe resections, and 0.72(0.32) in extratemporal resections. In comparison to other pipelines, RAMPS had higher dice similarities (N=62) (RAMPS:0.86, Epic-CHOP: 0.72, ResectVol: 0.72). We release a user-friendly, easy to use pipeline, RAMPS, open source for accurate delineation of resected tissue.

q-bio.NC↗

arXiv2025

From Bedside to Desktop: A Data Protocol for Normative Intracranial EEG and Abnormality Mapping

Normative mapping is a framework used to map population-level features of health-related variables. It is widely used in neuroscience research, but the literature lacks established protocols in modalities that do not support healthy control measurements, such as intracranial EEG (icEEG). An icEEG normative map would allow researchers to learn about population-level brain activity and enable comparison of individual data against these norms to identify abnormalities. Currently, no standardised guide exists for transforming clinical data into a normative, regional icEEG map. Papers often cite different software and numerous articles to summarise the lengthy method, making it laborious for other researchers to understand or apply the process. Our protocol seeks to remedy this gap by providing a dataflow guide and key decision points that summarise existing methods. This protocol is used heavily in published works from our own lab (twelve peer-reviewed journal publications). Briefly, we take as input, icEEG recordings and neuroimaging data from people with epilepsy who are undergoing evaluation for resective surgery. As final outputs, we obtain a normative icEEG map, comprising signal properties localised to brain regions. Optionally, we can also process new subjects through the same pipeline and obtain their z-scores (or centiles) in each brain region, for abnormality detection and localisation. To date, a single, cohesive, dataflow pipeline for generating normative icEEG maps, along with abnormality mapping, has not been created. We envisage that this dataflow guide will not only increase understanding and application of normative mapping methods, but will also improve the consistency and quality of studies in the field.

q-bio.NC↗

arXiv2024

Efficient Task Grouping Through Samplewise Optimisation Landscape Analysis

Shared training approaches, such as multi-task learning (MTL) and gradient-based meta-learning, are widely used in various machine learning applications, but they often suffer from negative transfer, leading to performance degradation in specific tasks. While several optimisation techniques have been developed to mitigate this issue for pre-selected task cohorts, identifying optimal task combinations for joint learning - known as task grouping - remains underexplored and computationally challenging due to the exponential growth in task combinations and the need for extensive training and evaluation cycles. This paper introduces an efficient task grouping framework designed to reduce these overwhelming computational demands of the existing methods. The proposed framework infers pairwise task similarities through a sample-wise optimisation landscape analysis, eliminating the need for the shared model training required to infer task similarities in existing methods. With task similarities acquired, a graph-based clustering algorithm is employed to pinpoint near-optimal task groups, providing an approximate yet efficient and effective solution to the originally NP-hard problem. Empirical assessments conducted on 8 different datasets highlight the effectiveness of the proposed framework, revealing a five-fold speed enhancement compared to previous state-of-the-art methods. Moreover, the framework consistently demonstrates comparable performance, confirming its remarkable efficiency and effectiveness in task grouping.

cs.LG↗