arXiv subjects

Donghun Lee

Publications and source records attributed to Donghun Lee.

At least 19 recordsLinked to original sources

arXiv2026

Dissect and Prune: Enhancing Robustness in AI-Generated Image Detection

While existing AI-generated image detectors report high performance, we identify that this is largely driven by a critical prediction asymmetry: a bias toward the real class that severely limits sensitivity to generated content, especially under standard post-processing operations such as compression and resizing. We hypothesize that this stems from the model's reliance on spurious features, distracting signals that obscure true generative artifacts. To address this, we propose DEAR (Dissect and Prune), which leverages inpainted images to identify and prune these interfering components. Specifically, we find that features strongly aligned to either inpainted or non-inpainted regions are less robust to post-processing. By measuring the alignment between channel activations and inpaint masks, DEAR removes features at both extremes, retaining only those that capture genuine generative artifacts. Experimental results demonstrate that our approach significantly enhances robustness against unseen generators and post-processing, effectively mitigating the prediction asymmetry. Our code is available at https://github.com/dahyedahye/dear.

cs.CV↗

arXiv2026

One-Point Contraction: Erasing Representational Separability toward Irreversible Deep Forgetting

Machine unlearning is usually evaluated by what the classifier outputs: forget-set accuracy, confidence, membership-inference scores. We show that this is not enough. Across 14 representative unlearning methods on CIFAR-10 and SVHN, a single linear map fitted on a held-out calibration set, with no access to the forgotten data, reverses the unlearning in seconds and recovers forget-set accuracy to within a few percent of the original model. Recovered features even support pixel-level reconstruction through a generic decoder. We call this diagnostic Feature Mapping Recovery (FM-recovery). The pattern it exposes is uniform: current unlearning methods do not erase information from the representation, they apply an invertible linear distortion that hides it from one particular prediction head. We propose One-Point Contraction (OPC), an unlearning objective that collapses forget-set features to the origin while leaving the retain-set geometry intact. We prove that this contraction is equivalent to driving the predictive distribution to maximum entropy, so the same mechanism delivers behavioral forgetting and representation-level erasure at once. Forgotten queries land in a region the network treats as out-of-distribution, and the gradient signal on those queries collapses along with their features. OPC is the only method in our benchmark that survives FM-recovery, resists relearning and gradient-inversion attacks, and decouples forget from retain features in entangled settings, all without sacrificing retain or test accuracy.

cs.LG↗

arXiv2026

IPPRO: Importance-based Pruning with PRojective Offset for Magnitude-indifferent Structural Pruning

Importance-based structured pruning overwhelmingly relies on filter magnitude. This proxy is fundamentally flawed: due to scale invariance, functionally identical filters can receive arbitrarily different importance scores under rescaling. We propose IPPRO (Importance-based Pruning with PROjective Offset), a scale-invariant pruning framework grounded in projective geometry. By embedding filters into real projective space ($\mathbb{RP}^N$), IPPRO resolves the singularity at the origin, placing all filters at an equal angular distance from the zero filter. We define PROscore, which captures functional importance by measuring a filter's angular displacement toward zero under a single gradient step (directional collapse). We further connect PROscore to exact $L_0$ relaxation, proving this one-shot criterion reliably predicts multi-step pruning dynamics. Extensive experiments across CNNs, Vision Transformers, and LLMs (e.g., ResNet, DeiT, LLaMA) demonstrate that IPPRO consistently outperforms existing methods, yielding particularly striking gains under high compression and no-fine-tuning regimes, IPPRO establishes a robust, architecture-agnostic paradigm for neural network compression.

cs.LG↗

arXiv2026

RIPPLE: Generating Multi-Channel Phase, Not Recovering It

Generative models synthesize magnitude spectra with high fidelity, while phase is delegated to a recovery module---Griffin--Lim, a vocoder, or a latent decoder---applied independently to each channel. For multi-channel waveforms this delegation is costly: the physical content of spatial audio and three-component seismograms lives in the phase relationships between channels, precisely what channel-independent recovery cannot produce. The cost is also invisible, since the magnitude-based metrics common to both fields barely move when inter-channel phase coherence collapses---so a pipeline can discard the physical information in its output while still scoring well. We argue that phase should be generated, not recovered, and present RIPPLE (Rectified Inter-channel Phase with Prior-based LEarning), which reinterprets Griffin--Lim as a phase **prior** rather than a final estimator: initialized from the source phase, this prior carries the inter-channel structure to be preserved, and a rectified flow refines it toward the target under an explicit inter-channel phase loss. Tested on first-order ambisonics environment transfer and seismic cross-station translation---two physically unrelated domains---RIPPLE outperforms recovery-based pipelines on the coherence metrics that downstream analyses consume. The seismic case is decisive: across architecturally distinct generators, per-channel recovery leaves S-wave polarization error near the $57.3^\circ$ random expectation, whereas learned phase reduces it to $33.8^\circ$.

cs.LG↗

arXiv2026

Prime Fourier Embeddings: A Principled Basis for Modular Arithmetic

Numbers have algebraic structure that standard neural embeddings often fail to expose. We introduce Prime Fourier Embeddings (PFE), which encode integers as prime-indexed (cos, sin) pairs derived from the harmonic analysis of Q, providing a pre-structured representation in which modular arithmetic reduces to selecting the relevant prime channel rather than discovering algebraic structure from scratch. We prove that any linear map equivariant with respect to the product group action on PFE must be block-diagonal with one independent block per prime -- a consequence of Schur's lemma applied to the resulting character decomposition. For square-free composite moduli, the Chinese Remainder Theorem predicts which prime channels are task-relevant. Both predictions are confirmed empirically: ablation studies show specialization ratios exceeding 500x between task-relevant and task-irrelevant channels, with perfect in-distribution test accuracy across all square-free composite moduli tested.

cs.LG↗

arXiv2026

Strain-Induced Detuning of a Dressed Nitrogen-Vacancy Qubit: Effective Two-Level Theory and Its Validity

The nitrogen-vacancy (NV) center in diamond can be operated as a microwave-dressed qubit. In the ideal two-level limit, its transition frequency is first-order insensitive to static magnetic fields, providing robustness against magnetic detuning noise. In practical diamond devices, however, residual transverse crystal strain mixes the $\ket{m_{s}=\pm1}$ spin sublevels and modifies the dressed qubit. In this study, we derive an analytical effective two-level model of a strained dressed NV qubit by perturbatively eliminating the far-detuned spectator state from the full three-level dressed Hamiltonian. We obtain closed-form expressions for the dressed-state splitting, the spin-locking mixing angle, and the longitudinal magnetic-field coupling. We show that transverse strain shifts the dressed-state resonance and tilts the spin-locking axis. These two effects restore a finite DC-field response and thereby quantify the loss of magnetic robustness. We demonstrate these features in simulated pulsed electron spin resonance spectra that incorporate rate-equation-based optical readout. We further derive exact validity criteria from the eigenvalues and spectator weights of the full three-level Hamiltonian. For practical use, we reduce these criteria to two controlled guidelines: the spectator-like branch must remain above the nominal upper dressed state, and its branch-specific admixture must remain small. A validity diagram over the axial-field--transverse-strain plane summarizes these approximate conditions and provides practical guidelines for designing dressed-NV sensing experiments.

quant-ph↗

arXiv2026

Numbers Already Carry Their Own Embeddings

We introduce Adelic operation-preserved embeddings (AOE), a training-free representation that captures both a number's real value and its modular (p-adic) signatures. This construction preserves additive and multiplicative structure by design, turning numerical input into embeddings that "speak in the language of mathematics." Unlike prior approaches that rely on task-specific retraining, AOE is plug-and-play and drops seamlessly into existing architectures. On algebraic combinatorics benchmarks, it delivers consistent gains including the first-ever perfect accuracy on the Weaving Pattern task-while suggesting a principled path forward for overcoming the long-standing "number problem" in AI.

cs.LG↗

arXiv2026

LFNO: Bridging Laplace and Fourier via Transient-Steady Decomposition

We introduce the Laplace-Fourier Neural Operator (LFNO), a unified framework for modeling dynamical systems across transient and steady-state regimes by integrating the spectral advantages of Laplace and Fourier Neural Operators. LFNO employs a dual-branch architecture that explicitly decomposes system dynamics into transient and steady-state components. We evaluate LFNO on nine benchmarks, including three ODE systems (Duffing, Lorenz, and Pendulum) and six PDE systems (Euler-Bernoulli beam, Heat, Reaction-diffusion, Brusselator, Burgers, and Navier-Stokes). LFNO significantly outperforms existing operators on ODE systems, where transient dynamics dominate, and consistently surpasses LNO while achieving performance competitive with FNO on PDE benchmarks. Furthermore, LFNO offers improved stability and physical interpretability through its component-wise decomposition. These results demonstrate that LFNO provides a robust and unified approach for learning complex dynamical systems across multiple temporal scales.

cs.LG↗

arXiv2026

Virchow: A Million-Slide Digital Pathology Foundation Model

The use of artificial intelligence to enable precision medicine and decision support systems through the analysis of pathology images has the potential to revolutionize the diagnosis and treatment of cancer. Such applications will depend on models' abilities to capture the diverse patterns observed in pathology images. To address this challenge, we present Virchow, a foundation model for computational pathology. Using self-supervised learning empowered by the DINOv2 algorithm, Virchow is a vision transformer model with 632 million parameters trained on 1.5 million hematoxylin and eosin stained whole slide images from diverse tissue and specimen types, which is orders of magnitude more data than previous works. The Virchow model enables the development of a pan-cancer detection system with 0.949 overall specimen-level AUC across 17 different cancer types, while also achieving 0.937 AUC on 7 rare cancer types. The Virchow model sets the state-of-the-art on the internal and external image tile level benchmarks and slide level biomarker prediction tasks. The gains in performance highlight the importance of training on massive pathology image datasets, suggesting scaling up the data and network architecture can improve the accuracy for many high-impact computational pathology applications where limited amounts of training data are available.

eess.IV↗

arXiv2026

ASAP: Attention Sink Anchored Pruning

Vision Transformers (ViTs) face severe computational bottlenecks due to the quadratic complexity of self-attention at high resolutions. Existing token reduction methods rely on local metrics - such as single-layer attention scores - that are inherently vulnerable to the attention sink phenomenon, where uninformative tokens are paradoxically preserved over salient foreground objects. We propose ASAP (Attention Sink Anchored Pruning), a training-free framework that recasts this sink as a feature. Modeling ViT information flow as a Lazy Random Walk, ASAP identifies the sink as a dominant accumulator of probability mass. By computing the diffusion distance to the sink within the cumulative transition matrix, ASAP partitions tokens via Radial Diffusion Clustering and compresses background redundancy through Transition Weight Pooling in a single shot. Extensive experiments across image, video, and vision-language tasks demonstrate ASAP outperforms state-of-the-art methods, accelerating throughput by up to 48% while maintaining - or even exceeding - baseline accuracy.

cs.LG↗

arXiv2026

The Weight Gram Matrix Captures Sequential Feature Linearization in Deep Networks

Understanding how deep neural networks learn representations remains a central challenge in machine learning theory. In this work, we propose a feature-centric framework for analyzing neural network training by relating weight updates to feature evolution. We introduce a simple identity, the Feature Learning Equation, which identifies the weight Gram matrix as the key object capturing feature dynamics. This enables us to interpret gradient descent as implicitly inducing a hypothetical evolution of features, whose covariance structure - termed the Virtual Covariance - characterizes how representations evolve during training. Building on this perspective, we introduce Target Linearity, a measure quantifying the linear alignment between features and targets. By analyzing the training and layer-wise dynamics, we show that deep networks learn to sequentially transform representations toward target-linear structure. This linearization perspective provides a unified interpretation of several empirical phenomena, including Neural Collapse and linear interpolation in generative models.

cs.LG↗

arXiv2025

Dissecting and Re-architecting 3D NAND Flash PIM Arrays for Efficient Single-Batch Token Generation in LLMs

The advancement of large language models has led to models with billions of parameters, significantly increasing memory and compute demands. Serving such models on conventional hardware is challenging due to limited DRAM capacity and high GPU costs. Thus, in this work, we propose offloading the single-batch token generation to a 3D NAND flash processing-in-memory (PIM) device, leveraging its high storage density to overcome the DRAM capacity wall. We explore 3D NAND flash configurations and present a re-architected PIM array with an H-tree network for optimal latency and cell density. Along with the well-chosen PIM array size, we develop operation tiling and mapping methods for LLM layers, achieving a 2.4x speedup over four RTX4090 with vLLM and comparable performance to four A100 with only 4.9% latency overhead. Our detailed area analysis reveals that the proposed 3D NAND flash PIM architecture can be integrated within a 4.98mm2 die area under the memory array, without extra area overhead.

cs.AR↗

arXiv2025

DiaTool-DPO: Multi-Turn Direct Preference Optimization for Tool-Augmented Large Language Models

Tool-Augmented Larage Language Models (TA-LLMs) have shown promise in real-world applications, but face challenges in handling incomplete queries and out-of-scope requests. While existing approaches rely mainly on Supervised Fine-Tuning with expert trajectories, we propose DiaTool-DPO, a novel method that enhances TA-LLM's dialogue capabilities through Direct Preference Optimization. We model TA-LLM interactions as a Markov Decision Process with 5 distinct dialogue states and categorize user queries into 3 types based on their state transition trajectories. We automatically construct paired trajectory datasets of correct and incorrect dialogue flows and introduce a specialized objective loss for dialogue control. Our comprehensive evaluation demonstrates that DiaTool-DPO approaches GPT-4o's performance (94.8% in information gathering, 91% in tool call rejection) with substantial improvements over baseline (44% and 9.6% respectively) while maintaining core functionality. Our approach opens new possibilities for developing TA-LLMs that can handle diverse real-world scenarios without requiring additional expert demonstrations or human labeling.

cs.CL↗

arXiv2025

Data-Driven Dimensional Synthesis of Diverse Planar Four-bar Function Generation Mechanisms via Direct Parameterization

Dimensional synthesis of planar four-bar mechanisms is a challenging inverse problem in kinematics, requiring the determination of mechanism dimensions from desired motion specifications. We propose a data-driven framework that bypasses traditional equation-solving and optimization by leveraging supervised learning. Our method combines a synthetic dataset, an LSTM-based neural network for handling sequential precision points, and a Mixture of Experts (MoE) architecture tailored to different linkage types. Each expert model is trained on type-specific data and guided by a type-specifying layer, enabling both single-type and multi-type synthesis. A novel simulation metric evaluates prediction quality by comparing desired and generated motions. Experiments show our approach produces accurate, defect-free linkages across various configurations. This enables intuitive and efficient mechanism design, even for non-expert users, and opens new possibilities for scalable and flexible synthesis in kinematic design.

cs.LG↗

arXiv2025

Catalyst: a Novel Regularizer for Structured Pruning with Auxiliary Extension of Parameter Space

Structured pruning aims to reduce the size and computational cost of deep neural networks by removing entire filters or channels. The traditional regularizers such as L1 or Group Lasso and its variants lead to magnitude-biased pruning decisions, such that the filters with small magnitudes are likely to be pruned. Also, they often entail pruning results with almost zero margin around pruning decision boundary, such that tiny perturbation in a filter magnitude can flip the pruning decision. In this paper, we identify the precise algebraic condition under which pruning operations preserve model performance, and use the condition to construct a novel regularizer defined in an extended parameter space via auxiliary catalyst variables. The proposed Catalyst regularization ensures fair pruning chance for each filters with theoretically provable zero bias to their magnitude and robust pruning behavior achieved by wide-margin bifurcation of magnitudes between the preserved and the pruned filters. The theoretical properties naturally lead to real-world effectiveness, as shown by empirical validations of Catalyst Pruning algorithm. Pruning results on various datasets and models are superior to state-of-the-art filter pruning methods, and at the same time confirm the predicted robust and fair pruning characteristics of Catalyst pruning.

cs.LG↗

arXiv2025

Suppression of spin bath and low-frequency noise for sub-MHz AC magnetometry based on double-dressed spin qubit in diamond

We experimentally demonstrate a protocol that effectively suppresses the qubit-bath interaction in diamond and enables detection of weak AC signals (below 1 MHz) with enhanced signal-to-noise ratio (SNR) up to SNR = 17. The method is based on AC magnetometry with single- and double-dressed states that are adiabatically transferred from the initial qubit states using concatenated continuous dynamical decoupling. This work paves a way toward sensitive detection of weakly coupled nuclear spins in low-field NMR experiments.

quant-ph↗

arXiv2025

Broadband Ground Motion Synthesis by Diffusion Model with Minimal Condition

Shock waves caused by earthquakes can be devastating. Generating realistic earthquake-caused ground motion waveforms help reducing losses in lives and properties, yet generative models for the task tend to generate subpar waveforms. We present High-fidelity Earthquake Groundmotion Generation System (HEGGS) and demonstrate its superior performance using earthquakes from North American, East Asian, and European regions. HEGGS exploits the intrinsic characteristics of earthquake dataset and learns the waveforms using an end-to-end differentiable generator containing conditional latent diffusion model and hi-fidelity waveform construction model. We show the learning efficiency of HEGGS by training it on a single GPU machine and validate its performance using earthquake databases from North America, East Asia, and Europe, using diverse criteria from waveform generation tasks and seismology. Once trained, HEGGS can generate three dimensional E-N-Z seismic waveforms with accurate P/S phase arrivals, envelope correlation, signal-to-noise ratio, GMPE analysis, frequency content analysis, and section plot analysis.

cs.LG↗

arXiv2025

Re-experiment Smart: a Novel Method to Enhance Data-driven Prediction of Mechanical Properties of Epoxy Polymers

Accurate prediction of polymer material properties through data-driven approaches greatly accelerates novel material development by reducing redundant experiments and trial-and-error processes. However, inevitable outliers in empirical measurements can severely skew machine learning results, leading to erroneous prediction models and suboptimal material designs. To address this limitation, we propose a novel approach to enhance dataset quality efficiently by integrating multi-algorithm outlier detection with selective re-experimentation of unreliable outlier cases. To validate the empirical effectiveness of the approach, we systematically construct a new dataset containing 701 measurements of three key mechanical properties: glass transition temperature ($T_g$), tan $δ$ peak, and crosslinking density ($v_{c}$). To demonstrate its general applicability, we report the performance improvements across multiple machine learning models, including Elastic Net, SVR, Random Forest, and TPOT, to predict the three key properties. Our method reliably reduces prediction error (RMSE) and significantly improves accuracy with minimal additional experimental work, requiring only about 5% of the dataset to be re-measured. These findings highlight the importance of data quality enhancement in achieving reliable machine learning applications in polymer science and present a scalable strategy for improving predictive reliability in materials science.

cond-mat.soft↗