A5074727691- Topological and Geometric Data Analysis
- Remote-Sensing Image Classification
- Cell Image Analysis Techniques
- Advanced Neuroimaging Techniques and Applications
- Machine Learning in Materials Science
- Neural Networks and Applications
- Sparse and Compressive Sensing Techniques
- Anomaly Detection Techniques and Applications
- Advanced Image Processing Techniques
- Infrared Target Detection Methodologies
- Advanced Image Fusion Techniques
- Clusterin in disease pathology
- Adversarial Robustness in Machine Learning
- Advanced MRI Techniques and Applications
- Metallurgy and Material Forming
- Metal Forming Simulation Techniques
- Adaptive optics and wavefront sensing
- Generative Adversarial Networks and Image Synthesis
- Image Processing Techniques and Applications
- Solar Radiation and Photovoltaics
- Neural Networks and Reservoir Computing
- Advanced Chemical Sensor Technologies
- Image and Signal Denoising Methods
- Advanced Graph Neural Networks
- Underwater Acoustics Research
Pacific Northwest National Laboratory
2018-2024
The University of Texas at El Paso
2021-2023
Colorado State University
2014-2023
Seattle University
2022
Union College
2022
Australian National University
2022
Australian Mathematical Sciences Institute
2022
Naval Research Laboratory Optical Sciences Division
2019
United States Naval Research Laboratory
2018-2019
Optical Sciences (United States)
2018-2019
Many datasets can be viewed as a noisy sampling of an underlying space, and tools from topological data analysis characterize this structure for the purpose knowledge discovery. One such tool is persistent homology, which provides multiscale description homological features within dataset. A useful representation information persistence diagram (PD). Efforts have been made to map PDs into spaces with additional valuable machine learning tasks. We convert PD finite-dimensional vector we call...
Abstract We use methods from computational algebraic topology to study functional brain networks in which nodes represent regions and weighted edges encode the similarity of magnetic resonance imaging (fMRI) time series each region. With these tools, allow one characterize topological invariants such as loops high-dimensional data, we are able gain understanding low-dimensional structures a way that complements traditional approaches based on pairwise interactions. In present paper,...
This paper describes the 2nd edition of ICML Topological Deep Learning Challenge that was hosted within 2024 ELLIS Workshop on Geometry-grounded Representation and Generative Modeling (GRaM). The challenge focused problem representing data in different discrete topological domains order to bridge gap between (TDL) other types structured datasets (e.g. point clouds, graphs). Specifically, participants were asked design implement liftings, i.e. mappings structures --like hypergraphs, or...
Abstract The research and development cycle of advanced manufacturing processes traditionally requires a large investment time resources. Experiments can be expensive are hence conducted on relatively small scales. This poses problems for typically data-hungry machine learning tools which could otherwise expedite the cycle. We build upon prior work by applying conditional generative adversarial networks (GANs) to scanning electron microscope (SEM) imagery from an emerging process,...
A new model for turbulence-corrupted imagery is proposed based on the theory of optimal mass transport. By describing relationship between photon density and phase traveling wave, combining it with a least action principle, suggests class methods approximately recovering solution flow created by turbulent atmosphere. Both coherent incoherent are used to validate compare other typically describe this type data. Given its superior performance in experimental data, algorithms variety...
New depth sensors, like the Microsoft Kinect, produce streams of human pose data. These discrete can be viewed as noisy samples an underlying continuous ideal curve that describes a trajectory through high-dimensional space. This paper introduces technique for generalized curvature analysis (GCA) determines features along which used to characterize change and segment motion. Tools are developed approximating curvatures at mean points in terms singular values local mean-centered data balls....
Symmetry is a fundamental tool in the exploration of broad range complex systems. In machine learning symmetry has been explored both models and data. this paper we seek to connect symmetries arising from architecture family with that family's internal representation We do by calculating set groups, which call intertwiner groups model. model's representations data through experiments probe similarities between hidden states across same architecture. Our work suggests network are propagated...
Traditional models for beam broadening include both a diffractive term, owing to the source aperture, and ‘wandering’ term that stems from refractive index variations in atmosphere. Here we derive novel depends on properties of atmospheric turbulence. The derivation rests transport formulation propagation problem whereby magnitude electric field is viewed as density fluid, moving flow driven by perturbations. Properties solutions are obtained using Lagrangian coordinates demonstrated be...
Atmospheric correction is the process for removing atmospheric effects from spectral data; a necessary step recovering salient properties. The complex interactions between atmosphere and light are dominated by absorbance scattering physics. Existing methods modeling typically rely on deep knowledge of relevant environmental conditions high-fidelity numerical simulations governing physics in order to obtain accurate estimates these effects. Additionally, existing approaches often require...
It is often said that a deep learning model "invariant" to some specific type of transformation. However, what meant by this statement strongly depends on the context in which it made. In paper we explore nature invariance and equivariance models with goal better understanding ways they actually capture these concepts formal level. We introduce family metrics allows us quantify properties way disentangles them from other such as loss or accuracy. use our understand two most popular methods...
We present two fully unsupervised deep learning approaches for hyperspectral anomaly detection. In one approach we formulate the detection problem as an adversarial game where a generator network learns distribution of background pixels comprising single image and output corresponding discriminator yields statistic. The other formulates statistic error between input pixel reconstruction that by autoencoder trained on image. Both methods leverage sub-sampling scheme allows training testing...
We introduce an extension that may be used to augment algorithms for the sparse decomposition of signals into a linear combination atoms drawn from dictionary such as those in support of, example, compressive sensing, k-sparse representation, and denoising. Our augmentation applied any reconstruction algorithm relies on selection sorting high-correlation during analysis or identification phase by generating "path" between two highest-correlation atoms. Here we investigate types path:...
We use methods from computational algebraic topology to study functional brain networks, in which nodes represent regions and weighted edges encode the similarity of fMRI time series each region. With these tools, allow one characterize topological invariants such as loops high-dimensional data, we are able gain understanding into low-dimensional structures networks a way that complements traditional approaches based on pairwise interactions. In present paper, persistent homology analyze...
There is a growing body of work that leverages features extracted via topological data analysis to train machine learning models. While this field, sometimes known as (TML), has seen some notable successes, an understanding how the process from differs raw still limited. In work, we begin address one component larger issue by asking whether model trained with learns internal representations are fundamentally different than those learned original data. To quantify ``different'', exploit two...
We introduce a path-augmentation step to the standard orthogonal matching pursuit algorithm. Our augmentation may be applied any algorithm that relies on selection and sorting of high-correlation atoms during an analysis or identification phase by generating "path" between two highest-correlation atoms. Here we investigate types path: linear combination (Euclidean geodesic) construction relying optimal transport map (2-Wasserstein geodesic). test our extension k-sparse reconstructions faces...
We present a framework for inferring an atmospheric transmission profile from spectral scene. This leverages lightweight, physics-based simulator that is automatically tuned - by virtue of autodifferentiation and differentiable programming to construct surrogate model the observed data. demonstrate utility methodology (i) performing correction, (ii) recasting data between various modalities (e.g. radiance reflectance at surface sensor), (iii) profiles, such as absorbing bands their relative...
Given the scale and complexity of forthcoming HSI data, producing labeled datasets at required to improve state-of-the-art performance is impractical prohibitively costly. Unsupervised pre-training algorithms have revolutionized deep learning for natural language processing computer vision by tapping into vast troves unlabeled but these advances seen little adoption in domain. We present some early results from self-supervised hyperspectral imagery using masked auto-encoders compare...