arXiv subjects

Arya Bangun

Publications and source records attributed to Arya Bangun.

15 recordsLinked to original sources

arXiv2026

RECAST: Model Reconstruction via Counterfactual-Aware Wasserstein Geometry under Limited Data

Counterfactual explanations (CFs) help understand machine learning models by identifying minimal input changes that would lead to alternative model outcomes. Recent work demonstrates their utility for reconstructing black-box models, enabling third-party auditing of opaque decision systems for fairness and accountability. Still, CF-based reconstruction may suffer from decision boundary shifts, overfitting, and restrictive assumptions requiring online query access to target platforms. We propose REconstruction via Counterfactual-Aware waSserstein opTimization (RECAST) under limited data and restricted access, a behavioral surrogate model based on Wasserstein barycentric prototypes. Our approach addresses decision boundary shifts by incorporating CFs as informative, though less representative, samples for both classes, maintaining high surrogate fidelity in low-sample regimes without requiring online access during reconstruction. To enhance fairness auditing, our method enables systematic group fairness diagnostics. Experiments on real-world datasets and various setups show that RECAST effectively achieves high fidelity and query efficiency, as well as stable results even when the access is limited and noisy.

cs.LG↗

arXiv2026

Least but not Last: Fine-tuning Intermediate Principal Components for Better Performance-Forgetting Trade-Offs

Low-Rank Adaptation (LoRA) methods have emerged as crucial techniques for adapting large pre-trained models to downstream tasks under computational and memory constraints. However, they face a fundamental challenge in balancing task-specific performance gains against catastrophic forgetting of pre-trained knowledge, where existing methods provide inconsistent recommendations. This paper presents a comprehensive analysis of the performance-forgetting trade-offs inherent in low-rank adaptation using principal components of weight matrices as initialization. Our investigation reveals that fine-tuning intermediate components leads to better balance and robustness to high learning rates than first (PiSSA) and last (MiLoRA) components in existing work. Building on these findings, we provide practical guidelines for initialization of LoRA methods to balance the performance-forgetting trade-off. In a thorough empirical study on a variety of computer vision and NLP tasks we confirm that these guidelines achieve high accuracy and reduced forgetting.

cs.LG↗

arXiv2025

Classifier Reconstruction Through Counterfactual-Aware Wasserstein Prototypes

Counterfactual explanations provide actionable insights by identifying minimal input changes required to achieve a desired model prediction. Beyond their interpretability benefits, counterfactuals can also be leveraged for model reconstruction, where a surrogate model is trained to replicate the behavior of a target model. In this work, we demonstrate that model reconstruction can be significantly improved by recognizing that counterfactuals, which typically lie close to the decision boundary, can serve as informative though less representative samples for both classes. This is particularly beneficial in settings with limited access to labeled data. We propose a method that integrates original data samples with counterfactuals to approximate class prototypes using the Wasserstein barycenter, thereby preserving the underlying distributional structure of each class. This approach enhances the quality of the surrogate model and mitigates the issue of decision boundary shift, which commonly arises when counterfactuals are naively treated as ordinary training instances. Empirical results across multiple datasets show that our method improves fidelity between the surrogate and target models, validating its effectiveness.

cs.LG↗

arXiv2025

Physics-Guided Diffusion Priors for Multi-Slice Reconstruction in Scientific Imaging

Accurate multi-slice reconstruction from limited measurement data is crucial to speed up the acquisition process in medical and scientific imaging. However, it remains challenging due to the ill-posed nature of the problem and the high computational and memory demands. We propose a framework that addresses these challenges by integrating partitioned diffusion priors with physics-based constraints. By doing so, we substantially reduce memory usage per GPU while preserving high reconstruction quality, outperforming both physics-only and full multi-slice reconstruction baselines for different modalities, namely Magnetic Resonance Imaging (MRI) and four-dimensional Scanning Transmission Electron Microscopy (4D-STEM). Additionally, we show that the proposed method improves in-distribution accuracy as well as strong generalization to out-of-distribution datasets.

eess.IV↗

arXiv2025

FlowTIE: Flow-based Transport of Intensity Equation for Phase Gradient Estimation from 4D-STEM Data

We introduce FlowTIE, a neural-network-based framework for phase reconstruction from 4D-Scanning Transmission Electron Microscopy (STEM) data, which integrates the Transport of Intensity Equation (TIE) with a flow-based representation of the phase gradient. This formulation allows the model to bridge data-driven learning with physics-based priors, improving robustness under dynamical scattering conditions for thick specimen. The validation on simulated datasets of crystalline materials, benchmarking to classical TIE and gradient-based optimization methods are presented. The results demonstrate that FlowTIE improves phase reconstruction accuracy, fast, and can be integrated with a thick specimen model, namely multislice method.

cs.LG↗

arXiv2025

1LoRA: Summation Compression for Very Low-Rank Adaptation

Parameter-Efficient Fine-Tuning (PEFT) methods have transformed the approach to fine-tuning large models for downstream tasks by enabling the adjustment of significantly fewer parameters than those in the original model matrices. In this work, we study the "very low rank regime", where we fine-tune the lowest amount of parameters per linear layer for each considered PEFT method. We propose 1LoRA (Summation Low-Rank Adaptation), a compute, parameter and memory efficient fine-tuning method which uses the feature sum as fixed compression and a single trainable vector as decompression. Differently from state-of-the-art PEFT methods like LoRA, VeRA, and the recent MoRA, 1LoRA uses fewer parameters per layer, reducing the memory footprint and the computational cost. We extensively evaluate our method against state-of-the-art PEFT methods on multiple fine-tuning tasks, and show that our method not only outperforms them, but is also more parameter, memory and computationally efficient. Moreover, thanks to its memory efficiency, 1LoRA allows to fine-tune more evenly across layers, instead of focusing on specific ones (e.g. attention layers), improving performance further.

cs.CV↗

arXiv2024

Eigenstructure Analysis of Bloch Wave and Multislice Formulations for Dynamical Scattering in Transmission Electron Microscopy

We investigate the eigenstructure of matrix formulations used for modeling scattering processes within materials in transmission electron microscopy. Dynamical scattering is crucial for describing the interaction between an electron wave and the material under investigation. Unlike the Bloch wave formulation, which defines the transmission function via the scattering matrix, the traditional multislice method is lacking a pure transmission function due to the entanglement of electron waves with the propagation function. To address this, we reformulate the multislice method into a matrix framework, which we refer to as the transmission matrix. This allows a direct comparison to the scattering matrix derived from Bloch waves in terms of their eigenstructures. Through theory, we demonstrate their equivalence with eigenvectors related by a two-dimensional Fourier matrix, given that the eigenvalue angles differ by modulo $2\pi n$ (integer $n$). We numerically verify our findings as well as demonstrate the application of the eigenstructure for the estimation of the mean inner potential.

cond-mat.mtrl-sci↗

arXiv2024

MRI Reconstruction with Regularized 3D Diffusion Model (R3DM)

Magnetic Resonance Imaging (MRI) is a powerful imaging technique widely used for visualizing structures within the human body and in other fields such as plant sciences. However, there is a demand to develop fast 3D-MRI reconstruction algorithms to show the fine structure of objects from under-sampled acquisition data, i.e., k-space data. This emphasizes the need for efficient solutions that can handle limited input while maintaining high-quality imaging. In contrast to previous methods only using 2D, we propose a 3D MRI reconstruction method that leverages a regularized 3D diffusion model combined with optimization method. By incorporating diffusion based priors, our method improves image quality, reduces noise, and enhances the overall fidelity of 3D MRI reconstructions. We conduct comprehensive experiments analysis on clinical and plant science MRI datasets. To evaluate the algorithm effectiveness for under-sampled k-space data, we also demonstrate its reconstruction performance with several undersampling patterns, as well as with in- and out-of-distribution pre-trained data. In experiments, we show that our method improves upon tested competitors.

eess.IV↗

arXiv2022

Wigner Distribution Deconvolution Adaptation for Live Ptychography Reconstruction

We propose a modification of Wigner Distribution Deconvolution (WDD) to support live processing ptychography. Live processing allows to reconstruct and display the specimen transfer function gradually while diffraction patterns are acquired. For this purpose we reformulate WDD and apply a dimensionality reduction technique that reduces memory consumption and increases processing speed. We show numerically that this approach maintains the reconstruction quality of specimen transfer functions as well as reduces computational complexity during acquisition processes. Although we only present the reconstruction for Scanning Transmission Electron Microscopy (STEM) datasets, in general, the live processing algorithm we present in this paper can be applied to real-time ptychographic reconstruction for different fields of application.

cs.IT↗

arXiv2022

Optimizing Sensing Matrices for Spherical Near-Field Antenna Measurements

In this article, we address the problem of reducing the number of required samples for Spherical Near-Field Antenna Measurements (SNF) by using Compressed Sensing (CS). A condition to ensure the numerical performance of sparse recovery algorithms is the design of a sensing matrix with low mutual coherence. Without fixing any part of the sampling pattern, we propose sampling points that minimize the mutual coherence of the respective sensing matrix by using augmented Lagrangian method. Numerical experiments show that the proposed sampling scheme yields a higher recovery success in terms of phase transition diagram when compared to other known sampling patterns, such as the spiral and Hammersley sampling schemes. Furthermore, we also demonstrate that the application of CS with an optimized sensing matrix requires fewer samples than classical approaches to reconstruct the Spherical Mode Coefficients (SMCs) and far-field pattern.

cs.IT↗

arXiv2022

Inverse Multislice Ptychography by Layer-wise Optimisation and Sparse Matrix Decomposition

We propose algorithms based on an optimisation method for inverse multislice ptychography in, e.g. electron microscopy. The multislice method is widely used to model the interaction between relativistic electrons and thick specimens. Since only the intensity of diffraction patterns can be recorded, the challenge in applying inverse multislice ptychography is to uniquely reconstruct the electrostatic potential in each slice up to some ambiguities. In this conceptual study, we show that a unique separation of atomic layers for simulated data is possible when considering a low acceleration voltage. We also introduce an adaptation for estimating the illuminating probe. For the sake of practical application, we finally present slice reconstructions using experimental 4D scanning transmission electron microscopy (STEM) data.

cs.IT↗

arXiv2021

Live processing of momentum-resolved STEM data for first moment imaging and ptychography

A reformulated implementation of single-sideband ptychography enables analysis and display of live detector data streams in 4D scanning transmission electron microscopy (STEM) using the LiberTEM open-source platform. This is combined with live first moment and further virtual STEM detector analysis. Processing of both real experimental and simulated data shows the characteristics of this method when data is processed progressively, as opposed to the usual offline processing of a complete dataset. In particular, the single side band method is compared to other techniques such as the enhanced ptychographic engine in order to ascertain its capability for structural imaging at increased specimen thickness. Qualitatively interpretable live results are obtained also if the sample is moved, or magnification is changed during the analysis. This allows live optimization of instrument as well as specimen parameters during the analysis. The methodology is especially expected to improve contrast- and dose-efficient in-situ imaging of weakly scattering specimens, where fast live feedback during the experiment is required.

physics.data-an↗

arXiv2020

Tight bounds on the mutual coherence of sensing matrices for Wigner D-functions on regular grids

Many practical sampling patterns for function approximation on the rotation group utilizes regular samples on the parameter axes. In this paper, we relate the mutual coherence analysis for sensing matrices that correspond to a class of regular patterns to angular momentum analysis in quantum mechanics and provide simple lower bounds for it. The products of Wigner d-functions, which appear in coherence analysis, arise in angular momentum analysis in quantum mechanics. We first represent the product as a linear combination of a single Wigner d-function and angular momentum coefficients, otherwise known as the Wigner 3j symbols. Using combinatorial identities, we show that under certain conditions on the bandwidth and number of samples, the inner product of the columns of the sensing matrix at zero orders, which is equal to the inner product of two Legendre polynomials, dominates the mutual coherence term and fixes a lower bound for it. In other words, for a class of regular sampling patterns, we provide a lower bound for the inner product of the columns of the sensing matrix that can be analytically computed. We verify numerically our theoretical results and show that the lower bound for the mutual coherence is larger than Welch bound. Besides, we provide algorithms that can achieve the lower bound for spherical harmonics.

quant-ph↗

arXiv2020

Sensing Matrix Design and Sparse Recovery on the Sphere and the Rotation Group

In this paper, {the goal is to design deterministic sampling patterns on the sphere and the rotation group} and, thereby, construct sensing matrices for sparse recovery of band-limited functions. It is first shown that random sensing matrices, which consists of random samples of Wigner D-functions, satisfy the Restricted Isometry Property (RIP) with proper preconditioning and can be used for sparse recovery on the rotation group. The mutual coherence, however, is used to assess the performance of deterministic and regular sensing matrices. We show that many of widely used regular sampling patterns yield sensing matrices with the worst possible mutual coherence, and therefore are undesirable for sparse recovery. Using tools from angular momentum analysis in quantum mechanics, we provide a new expression for the mutual coherence, which encourages the use of regular elevation samples. We construct low coherence deterministic matrices by fixing the regular samples on the elevation and minimizing the mutual coherence over the azimuth-polarization choice. It is shown that once the elevation sampling is fixed, the mutual coherence has a lower bound that depends only on the elevation samples. This lower bound, however, can be achieved for spherical harmonics, which leads to new sensing matrices with better coherence than other representative regular sampling patterns. This is reflected as well in our numerical experiments where our proposed sampling patterns perfectly match the phase transition of random sampling patterns.

cs.IT↗

arXiv2016

Sparse recovery in Wigner-D basis expansion

We are concerned with the recovery of $s-$sparse Wigner-D expansions in terms of $N$ Wigner-D functions. Considered as a generalization of spherical harmonics, Wigner-D functions are eigenfunctions of Laplace-Beltrami operator and form an orthonormal system. However, since they are not uniformly bounded, the existing results on BOS do not apply. Using previously introduced preconditioning technique, a new orthonormal and bounded system is obtained for which RIP property can be established. We show that the number of sufficient samples for sparse recovery scales with ${N}^{1/6} \,s\, \log^3(s) \,\log(N)$. The phase transition diagram for this problem is also presented. We will also discuss the application of our results in the spherical near-field antenna measurement.

cs.IT↗