Search Results (11)

Improving robustness under distributional shifts remains a central challenge in machine learning. Although Sharpness-Aware Minimization (SAM) has proven effective in finding flatter minima for better generalization, it overlooks the heterogeneity in sharpness across different subpopulations, which can exacerbate performance gaps for minority or vulnerable groups. To address this challenge, we propose Group-gap Guided SAM (G²-SAM), a new optimization framework that promotes distributional robustness by steering flatness-seeking directions according to intergroup loss disparities. Our method estimates group-wise sharpness and adaptively refines perturbation strategies to minimize the worst-group loss while preserving model consistency. Through comprehensive experiments across various datasets, we show that G²-SAM achieves superior Worst-Group Accuracy and robustness, outperforming previous baselines. These findings highlight the importance of addressing group-specific geometry in the loss landscape to build more reliable and equitable neural networks. Full article

Federated Learning (FL) is a promising distributed machine learning framework that emphasizes privacy protection. However, inconsistencies between local optimization objectives and the global objective, commonly referred to as client drift, primarily arise due to non-independently and identically distributed (Non-IID) data, multiple local training steps, and partial client participation in training. The majority of current research tackling this challenge is mainly based on the empirical risk minimization (ERM) principle, while giving little consideration to the connection between the global loss landscape and generalization capability. This study proposes FedGAM, an innovative FL algorithm that incorporates Gradient Norm-Aware Minimization (GAM) to efficiently search for a local flat landscape. FedGAM specifically modifies the client model training objective to simultaneously minimize the loss value and first-order flatness, thereby seeking flat minima. To directly smooth the global flatness, we propose the more significant FedGAM-CV, which employs control variables to correct local updates, guiding each client to train models in a globally flat direction. Experiments on three datasets (CIFAR-10, MNIST, and FashionMNIST) demonstrate that our proposed algorithms outperform existing FL baselines, effectively finding flat minima and addressing the client drift problem. Full article

The inverse finite element method is a technique that can be used for material model parameter characterization. The literature shows that this approach may get caught in the local minima of the design space. These local minimum solutions often fit the material test data with small errors and are often mistaken for the optimal solution. The problem with these sub-optimal solutions becomes apparent when applied to different loading conditions where significant errors can be witnessed. The research of this paper presents a new method that resolves this issue for Mooney–Rivlin and builds on a previous paper that used flat planes, referred to as hyperplanes, to map the error functions, isolating the unique optimal solution. The new method alternatively uses a constrained optimization approach, utilizing equality constraints to evaluate the error functions. As a result, the design space’s curvature is taken into account, which significantly reduces the amount of variation between predicted parameters from a maximum of $1.934$% in the previous paper down to $0.1882$% in the results presented here. The results of this study demonstrate that the new method not only isolates the unique optimal solution but also drastically reduces the variation in the predicted parameters. The paper concludes that the presented new characterization method significantly contributes to the existing literature. Full article

Molecular crystals have shallow potential energy landscapes, with multiple local minima separated by very small differences in total energy. Predicting molecular packing and molecular conformation in the crystal generally requires ab initio methods of high accuracy, especially when polymorphs are involved. We used dispersion-corrected density functional theory (DFT-D) to assess the capabilities of an evolutionary algorithm (EA) for the crystal structure prediction (CSP) of well-known but challenging high-energy molecular crystals (HMX, RDX, CL-20, and FOX-7). While providing the EA with the experimental conformation of the molecule quickly re-discovers the experimental packing, it is more realistic to start instead from a naïve, flat, or neutral initial conformation, which reflects the limited experimental knowledge we generally have in the computational design of molecular crystals. By doing so, and using fully flexible molecules in fully variable unit cells, we show that the experimental structures can be predicted in fewer than 20 generations. Nonetheless, one must be aware that some molecular crystals have naturally hindered evolutions, requiring as many attempts as there are space groups of interest to predict their structures, and some may require the accuracy of all-electron calculations to discriminate between closely ranked structures. To save resources in this computationally demanding process, we showed that a hybrid xTB/DFT-D approach could be considered in a subsequent study to push the limits of CSP beyond 200+ atoms and for cocrystals. Full article

Despite its effectiveness in determining breakage function parameters (BFPs) for quantifying breakage characteristics in mineral grinding processes, the back-calculation method has limitations owing to the uncertainty regarding the distribution of the error function. In this work, using Korean uranium and molybdenum ores, we show that the limitation can be overcome by searching over a wide range of initial values based on the conjugate gradient method. We also visualized the distribution of the sum of squares of the error in the two-dimensional parameter space. The results showed that the error function was strictly convex, and the main problem in the back-calculation of the breakage functions was the flat surface of the objective function rather than the occurrence of local minima. Based on our results, we inferred that the flat surface problem could be significantly mitigated by searching over a wide range of initial values. Back-calculation using a wide range of initial values yields BFPs similar to those obtained from single-sized-feed breakage tests (SSFBTs) up to four-dimensional parameter spaces. Therefore, by searching over a wide range of initial values, the feasibility of the back-calculation approach can be significantly improved with a minimum number of SSFBTs. Full article

A wide variety of neurodegenerative diseases are characterized by the accumulation of protein aggregates in intraneuronal or extraneuronal brain regions. In Alzheimer’s disease (AD), the extracellular aggregates originate from amyloid-β proteins, while the intracellular aggregates are formed from microtubule-binding tau proteins. The amyloid forming peptide sequences in the amyloid-β peptides and tau proteins are responsible for aggregate formation. Experimental studies have until the date reported many of such amyloid forming peptide sequences in different proteins, however, there is still limited molecular level understanding about their tendency to form aggregates. In this study, we employed umbrella sampling simulations and subsequent electronic structure theory calculations in order to estimate the energy profiles for interconversion of the helix to β-sheet like secondary structures of sequences from amyloid-β protein (KLVFFA) and tau protein (QVEVKSEKLD and VQIVYKPVD). The study also included a poly-alanine sequence as a reference system. The calculated force-field based free energy profiles predicted a flat minimum for monomers of sequences from amyloid and tau proteins corresponding to an α-helix like secondary structure. For the parallel and anti-parallel dimer of KLVFFA, double well potentials were obtained with the minima corresponding to α-helix and β-sheet like secondary structures. A similar double well-like potential has been found for dimeric forms for the sequences from tau fibril. Complementary semi-empirical and density functional theory calculations displayed similar trends, validating the force-field based free energy profiles obtained for these systems. Full article

The bacterial foraging optimization (BFO) algorithm successfully searches for an optimal path from start to finish in the presence of obstacles over a flat surface map. However, the algorithm suffers from getting stuck in the local minima whenever non-circular obstacles are encountered. The retrieval from the local minima is crucial, as otherwise, it can cause the failure of the whole task. This research proposes an improved version of BFO called robust bacterial foraging (RBF), which can effectively avoid obstacles, both of circular and non-circular shape, without falling into the local minima. The virtual obstacles are generated in the local minima, causing the robot to retract and regenerate a safe path. The proposed method is easily extendable to multiple robots that can coordinate with each other. The information related to the virtual obstacles is shared with the whole swarm, so that they can escape the same local minima to save time and energy. To test the effectiveness of the proposed algorithm, a comparison is made against the existing BFO algorithm. Through the results, it was witnessed that the proposed approach successfully recovered from the local minima, whereas the BFO got stuck. Full article

In this work, we demonstrate spectral phase-shift interferometry operating in the near-infrared wavelength range for refractive index (RI) monitoring of fluidic samples in micro-capillaries. A detailed theoretical model was developed to calculate the phase-sensitive spectral reflectivity when low-cost rectangular glass micro-capillaries, filled with samples with different refractive indices, are placed at the end of the measurment arm of a Michelson interferometer. From the phase-sensitive spectral reflectivity, we recovered the cosine-shaped interferometric signal as a function of the wavelength, as well as its dependence on the sample RI. Using the readout radiation provided by a 40-nm wideband light source with a flat emission spectrum centered at 1.55 µm and a 2 × 1 fiberoptic coupler on the common input-output optical path, experimental results were found to be in good agreement with the expected theoretical behavior. The shift of the micro-capillary optical resonances, induced by RI variations in the filling fluids (comparing saline solution with respect to distilled water, and isopropanol with respect to ethanol) were clearly detected by monitoring the positions of steep phase jumps in the cosine-shaped interferometric signal recorded as a function of the wavelength. By adding a few optical components to the instrumental configuration previously demonstrated for the spectral amplitude detection of resonances, we achieved phase-sensitive detection of the wavelength positions of the resonances as a function of the filling fluid RI. The main advantage consists of recovering RI variations by detecting the wavelength shift of “sharp peaks”, with any amplitude above a threshold in the interferometric signal derivative, instead of “wide minima” in the reflected power spectra, which are more easily affected by uncertainties due to amplitude fluctuations. Full article

The paper studies the modified theory of induced gravity (MTIG). The solutions of the MTIG equations contain two branches (stages): Einstein (ES) and “restructuring” (RS). Previously, solutions were found that the values of such parameters as the “Hubble parameter”, gravitational and cosmological “constants” at the RS stage, fluctuate near monotonously developing mean values. This article gives MTIG equations with arbitrary potential. Solutions of the equations of geodesic curves are investigated for the case of centrally symmetric space and quadratic potential at the RS stage. The oscillatory nature of the solutions leads to the appearance of a gravitational potential containing a spectrum of minima, as well as to antigravity, which is expressed by acceleration directed from the center. Such solutions lead to the distribution of the potential of the gravitational field creating an additional mass effect at large distances and are well suited for modeling the effect of dark matter in galaxies. The solutions of the equation of geodesic lines are obtained and analyzed. We found that the transition from flat asymptotics to oscillatory asymptotics at large distances from the center with a combination of the presence of antigravity zones leads to a rich variety of shapes and dynamics of geodesic curves and to the formation of complex structures. Full article

We hypothesize that machining grooves and ridges into the surface of radiata pine deck boards will change the pattern of strains that develop when profiled boards are exposed to wetting and drying. Two wavy profiles were tested, and flat unprofiled boards acted as controls. Full-field surface strain data was collected using digital image correlation. Strains varied across the surface of both flat and profiled boards during wetting and drying. Profiling fundamentally changed surface strain patterns; strain maxima and minima developed in the profile ridges and grooves during wetting, respectively, but this pattern of strains reversed during drying. Such a pronounced reversal of strains was not observed when flat boards were exposed to wetting and drying, although there was a shift towards negative strains when flat boards were dried. We conclude that profiling changes surface strain distribution in deck boards exposed to wetting and drying, and causes high strains to develop in the grooves of profiled boards. These findings help explain why checks in profiled deck boards are mainly confined to profile grooves where they are difficult to see, and the commercial success of profiling at reducing the negative effects of checking on the appearance of wood decking. Full article

Most existing ICP (Iterative Closet Point)-based 3D ear recognition approaches resort to the coarse-to-fine ICP algorithms to match 3D ear models. With such an approach, the gallery-probe pairs are coarsely aligned based on a few local feature points and then finely matched using the original ear point cloud. However, such an approach ignores the fact that not all the points in the coarsely segmented ear data make positive contributions to recognition. As such, the coarsely segmented ear data which contains a lot of redundant and noisy data could lead to a mismatch in the recognition scenario. Additionally, the fine ICP matching can easily trap in local minima without the constraint of local features. In this paper, an efficient and fully automatic 3D ear recognition system is proposed to address these issues. The system describes the 3D ear surface with a local feature—the Local Surface Variation (LSV), which is responsive to the concave and convex areas of the surface. Instead of being used to extract discrete key points, the LSV descriptor is utilized to eliminate redundancy flat non-ear data and get normalized and refined ear data. At the stage of recognition, only one-step modified iterative closest points using local surface variation (ICP-LSV) algorithm is proposed, which provides additional local feature information to the procedure of ear recognition to enhance both the matching accuracy and computational efficiency. On an Inter^®Xeon^®W3550, 3.07 GHz work station (DELL T3500, Beijing, China), the authors were able to extract features from a probe ear in 2.32 s match the ear with a gallery ear in 0.10 s using the method outlined in this paper. The proposed algorithm achieves rank-one recognition rate of 100% on the Chinese Academy of Sciences’ Institute of Automation 3D Face database (CASIA-3D FaceV1, CASIA, Beijing, China, 2004) and 98.55% with 2.3% equal error rate (EER) on the Collection J2 of University of Notre Dame Biometrics Database (UND-J2, University of Notre Dame, South Bend, IN, USA, between 2003 and 2005). Full article