Resource
2025 EN
Hasan Berkay Abdioglu · Rana Gursoy · Yagmur Isik
+8 more
This study investigates the application of Super-Resolution techniques inholographic microscopy to enhance quantitative phase imaging. An off-axisMach-Zehnder interferometric setup was employed to capture interferograms. Thestudy evaluates two Super-Resolution models, RCAN and Real-ESRGAN, for theireffectiveness in reconstructing high-resolution interferograms from amicroparticle-based dataset. The models were assessed using two primaryapproaches: image-based analysis for structural detail enhancement andmorphological evaluation for maintaining sample integrity and phase mapaccuracy. The results demonstrate that RCAN achieves superior numericalprecision, making it ideal for applications requiring highly accurate phase mapreconstruction, while Real-ESRGAN enhances visual quality and structuralcoherence, making it suitable for visualization-focused applications. Thisstudy highlights the potential of Super-Resolution models in overcomingdiffraction-imposed resolution limitations in holographic microscopy, openingthe way for improved imaging techniques in biomedical diagnostics, materialsscience, and other high-precision fields.
Resource
2025 EN
Kerem Zaman · Shashank Srivastava
Large Language Models (LLMs) offer natural language explanations as analternative to feature attribution methods for model interpretability. However,despite their plausibility, they may not reflect the model's internal reasoningfaithfully, which is crucial for understanding the model's true decision-makingprocesses. Although several faithfulness metrics have been proposed, a unifiedevaluation framework remains absent. To address this gap, we present CausalDiagnosticity, a framework to evaluate faithfulness metrics for naturallanguage explanations. Our framework employs the concept of causaldiagnosticity, and uses model-editing methods to generate faithful-unfaithfulexplanation pairs. Our benchmark includes four tasks: fact-checking, analogy,object counting, and multi-hop reasoning. We evaluate a variety of faithfulnessmetrics, including post-hoc explanation and chain-of-thought-based methods. Wefind that all tested faithfulness metrics often fail to surpass a randombaseline. Our work underscores the need for improved metrics and more reliableinterpretability methods in LLMs.
Resource
2025 EN
M Mahmudul Hasan Sajeeb · Navid Anjum Aadit · Shuvro Chowdhury
+7 more
In recent years, hardware implementations of Ising machines have emerged as aviable alternative to quantum computing for solving hard optimization problemsamong other applications. Unlike quantum hardware, dense connectivity can beachieved in classical systems. However, we show that dense connectivity leadsto severe frequency slowdowns and interconnect congestion scaling unfavorablywith system sizes. As a scalable solution, we propose a systematicsparsification method for dense graphs by introducing copy nodes to limit thenumber of neighbors per graph node. In addition to solving interconnectcongestion, this approach enables constant frequency scaling where all spins ina network can be updated in constant time. On the other hand, sparsificationintroduces new difficulties, such as constraint-breaking between copied spinsand increased convergence times to solve optimization problems, especially ifexact ground states are sought. Relaxing the exact solution requirements, wefind that the overheads in convergence times are milder. We demonstrate theseideas by designing probabilistic bit Ising machines using ASAP7 (a predictive7nm FinFET technology model) process design kits as well as Field ProgrammableGate Array (FPGA)-based implementations. Finally, we show how formulatingproblems in naturally sparse networks (e.g., by invertible logic) sidestepschallenges introduced by sparsification methods. Our results are applicable toa broad family of Ising machines using different hardware implementations.
Resource
2025 EN
Shuvro Chowdhury · Navid Anjum Aadit · Andrea Grimaldi
+12 more
Recent demonstrations on specialized benchmarks have reignited excitement forquantum computers, yet whether they can deliver an advantage for practicalreal-world problems remains an open question. Here, we show that probabilisticcomputers (p-computers) when co-designed with hardware to implement powerfulMonte Carlo algorithms can surpass state-of-the-art quantum annealers [Kinget al., Nature (2023)] in solving certain hard optimization problems. Wefocus on two key algorithms: discrete-time simulated quantum annealing (DT-SQA)and adaptive parallel tempering (APT), both applied to 3D spin glasses. ForDT-SQA, we find that increasing the number of replicas improves residual energyscaling, while parallelizing fewer replicas across independent runs alsoachieves comparable scaling. Both strategies align with the theoreticalexpectations from extreme value theory. In addition, APT outperforms DT-SQAwhen supported by non-local isoenergetic cluster moves. Finite-size scalinganalysis suggests a universal behavior that explains the superior performanceof APT over both DT-SQA and quantum annealing. We show that these algorithmsare readily implementable in modern hardware thanks to the mature semiconductortechnology. Unlike software simulations, replicas can be monolithically housedon a single chip and a large number of spins can be updated in parallel andasynchronously, similar to a quantum annealer. We project that custom FieldProgrammable Gate Arrays (FPGA) or specialized chips leveraging massiveparallelism can further accelerate these algorithms by orders of magnitude,while drastically improving energy efficiency. Our results raise the bar for apractical quantum advantage in optimization and present p-computers asscalable, energy-efficient hardware for real-world optimization problems.
Resource
2025 EN
Mehmet Kerem Turkcan · Mattia Ballo · Filippo Filicori
+1 more
We introduce specialized diffusion-based generative models that capture thespatiotemporal dynamics of fine-grained robotic surgical sub-stitch actionsthrough supervised learning on annotated laparoscopic surgery footage. Theproposed models form a foundation for data-driven world models capable ofsimulating the biomechanical interactions and procedural dynamics of surgicalsuturing with high temporal fidelity. Annotating a dataset of $\sim2K$ clipsextracted from simulation videos, we categorize surgical actions intofine-grained sub-stitch classes including ideal and non-ideal executions ofneedle positioning, targeting, driving, and withdrawal. We fine-tune twostate-of-the-art video diffusion models, LTX-Video and HunyuanVideo, togenerate high-fidelity surgical action sequences at $\ge$768x512 resolution and$\ge$49 frames. For training our models, we explore both Low-Rank Adaptation(LoRA) and full-model fine-tuning approaches. Our experimental resultsdemonstrate that these world models can effectively capture the dynamics ofsuturing, potentially enabling improved training simulators, surgical skillassessment tools, and autonomous surgical systems. The models also display thecapability to differentiate between ideal and non-ideal technique execution,providing a foundation for building surgical training and evaluation systems.We release our models for testing and as a foundation for future research.Project Page: https://mkturkcan.github.io/suturingmodels/
Resource
2025 EN
M. Kerem Aydin · Yi-Chun Hung · Jaclyn Pytlarz
+2 more
Hyperspectral cameras face harsh trade-offs between spatial, spectral, andtemporal resolution in an inherently low-photon regime. Computational imagingsystems break through these trade-offs with compressive sensing, but requirecomplex optics and/or extensive compute. We present Spectrum from Defocus(SfD), a chromatic focal sweep method that recovers state-of-the-arthyperspectral images with a small system of off-the-shelf optics and < 1 secondof compute. Our camera uses two lenses and a grayscale sensor to preservenearly all incident light in a chromatically-aberrated focal stack. Ourphysics-based iterative algorithm efficiently demixes, deconvolves, anddenoises the blurry grayscale focal stack into a sharp spectral image. Thecombination of photon efficiency, optical simplicity, and physical modelingmakes SfD a promising solution for fast, compact, interpretable hyperspectralimaging.
Resource
2025 EN
Kerem Bozkurt · Christoph Lohrmann · Felix Weinhardt
+5 more
Biofilms exposed to flow experience shear stress, which leads to acompetitive interaction between the growth and development of a biofilm andshearing. In this study, Pseudonomas fluorescene biofilm was grown in amicrofluidic channel and exposed to forced flow of an aqueous solution ofvariable velocity. It can be observed that under certain conditionspreferential flow paths form with a dynamic, but quasi-steady state interactionof growth, detachment, and re-attachment. We find that the regimes forpreferential flow path development are determined by nutrient availability andthe ratio of shear stress versus the biofilm's ability to resist shear forces.The intermittent regime of flow paths is mainly driven by the supply withnutrients, which we confirm by comparison with a numerical model based oncoarse-grained molecular dynamics and Lattice Boltzmann hydrodynamics.
Resource
2025 EN
Jinesh Jhonsa · William Whitehead · David McCarthy
+3 more
This paper demonstrates a probabilistic bit physics inspired solver with 440spins configured in a Chimera graph, occupying an area of 0.44 mm^2. Areaefficiency is maximized through a current-mode implementation of the neuronupdate circuit, standard cell design for analog blocks pitch-matched to digitalblocks, and a shared power supply for both digital and analog components.Process variation related mismatches introduced by this approach areeffectively mitigated using a hardware aware contrastive divergence algorithmduring training. We validate the chip's ability to perform probabilisticcomputing tasks such as modeling logic gates and full adders, as well asoptimization tasks such as MaxCut, demonstrating its potential for AI andmachine learning applications.
Journals
2024 EN
Kaya Kerem · Kravberg Alexander · Scarpellini Claudia
+3 more
“hot flow in the coldchanges shape in a dead worldcomes matter to life” Programmable matter that allows free shape transfiguration and locomotion on command promises ubiquitous access to objects or functions of interest. Current approaches for the autonomous reshaping of solid objects (smart materials, soft actuators, modular robotics) are limited in spatial resolution and shape. Solid‐liquid phase change pumping as a mechanism for the contactless transfiguring and locomotion of solid objects is introduced. Thin objects are deformed into any intended shape with sub‐millimeter resolution and the ability to freely change their topology is demonstrated, including adding or removing holes, splitting and merging. The unique locomotion of objects through millimeter‐sized constrictions narrower than their body size is demonstrated, followed by restoring the original shape. This approach opens up avenues for developing autonomous programmable matter with free shape transfiguration.
Journals
2024 EN
Artuk Kerem · Turkay Deniz · Mensi Mounir D.
+10 more
Abstract The primary performance limitation in inverted perovskite‐based solar cells is the interface between the fullerene‐based electron transport layers and the perovskite. Atomic layer deposited thin aluminum oxide (AlO X ) interlayers that reduce nonradiative recombination at the perovskite/C 60 interface are developed, resulting in >60 millivolts improvement in open‐circuit voltage and 1% absolute improvement in power conversion efficiency. Surface‐sensitive characterizations indicate the presence of a thin, conformally deposited AlO x layer, functioning as a passivating contact. These interlayers work universally using different lead‐halide–based absorbers with different compositions where the 1.55 electron volts bandgap single junction devices reach >23% power conversion efficiency. A reduction of metallic Pb 0 is found and the compact layer prevents in‐ and egress of volatile species, synergistically improving the stability. AlO X ‐modified wide‐bandgap perovskite absorbers as a top cell in a monolithic perovskite–silicon tandem enable a certified power conversion efficiency of 29.9% and open‐circuit voltages above 1.92 volts for 1.17 square centimeters device area.