Juan Castorena

Geigh Zollicoffer, Minh Vu, Ben Nebgen et al.

This work presents an information-theoretic examination of diffusion-based purification methods, the state-of-the-art adversarial defenses that utilize diffusion models to remove malicious perturbations in adversarial examples. By theoretically characterizing the inherent purification errors associated with the Markov-based diffusion purifications, we introduce LoRID, a novel Low-Rank Iterative Diffusion purification method designed to remove adversarial perturbation with low intrinsic purification errors. LoRID centers around a multi-stage purification process that leverages multiple rounds of diffusion-denoising loops at the early time-steps of the diffusion models, and the integration of Tucker decomposition, an extension of matrix factorization, to remove adversarial noise at high-noise regimes. Consequently, LoRID increases the effective diffusion time-steps and overcomes strong adversarial attacks, achieving superior robustness performance in CIFAR-10/100, CelebA-HQ, and ImageNet datasets under both white-box and black-box settings.

Minh Vu, Ben Nebgen, Erik Skau et al.

As Machine Learning (ML) applications rapidly grow, concerns about adversarial attacks compromising their reliability have gained significant attention. One unsupervised ML method known for its resilience to such attacks is Non-negative Matrix Factorization (NMF), an algorithm that decomposes input data into lower-dimensional latent features. However, the introduction of powerful computational tools such as Pytorch enables the computation of gradients of the latent features with respect to the original data, raising concerns about NMF's reliability. Interestingly, naively deriving the adversarial loss for NMF as in the case of ML would result in the reconstruction loss, which can be shown theoretically to be an ineffective attacking objective. In this work, we introduce a novel class of attacks in NMF termed Latent Feature Attacks (LaFA), which aim to manipulate the latent features produced by the NMF process. Our method utilizes the Feature Error (FE) loss directly on the latent features. By employing FE loss, we generate perturbations in the original data that significantly affect the extracted latent features, revealing vulnerabilities akin to those found in other ML techniques. To handle large peak-memory overhead from gradient back-propagation in FE attacks, we develop a method based on implicit differentiation which enables their scaling to larger datasets. We validate NMF vulnerabilities and FE attacks effectiveness through extensive experiments on synthetic and real-world data.

Juan Castorena, E. Louise Loudermilk, Scott Pokswinski et al.

The 3D structure of living and non-living components in ecosystems plays a critical role in determining ecological processes and feedbacks from both natural and human-driven disturbances. Anticipating the effects of wildfire, drought, disease, or atmospheric deposition depends on accurate characterization of 3D vegetation structure, yet widespread measurement remains prohibitively expensive and often infeasible. We introduce ForestGen3D, a novel generative modeling framework that synthesizes high-fidelity 3D forest structure using only aerial LiDAR (ALS) inputs. ForestGen3D is based on conditional denoising diffusion probabilistic models (DDPMs) trained on co-registered ALS/TLS (terrestrial LiDAR) data. The model learns to generate TLS-like 3D point clouds conditioned on sparse ALS observations, effectively reconstructing occluded sub-canopy detail at scale. To ensure ecological plausibility, we introduce a geometric containment prior based on the convex hull of ALS observations and provide theoretical and empirical guarantees that generated structures remain spatially consistent. We evaluate ForestGen3D at tree, plot, and landscape scales using real-world data from mixed conifer ecosystems, and show that it produces high-fidelity reconstructions that closely match TLS references in terms of geometric similarity and biophysical metrics, such as tree height, DBH, crown diameter and crown volume. Additionally, we demonstrate that the containment property can serve as a practical proxy for generation quality in settings where TLS ground truth is unavailable. Our results position ForestGen3D as a scalable tool for ecological modeling, wildfire simulation, and structural fuel characterization in ALS-only environments.

Juan Castorena

This work leverages neural radiance fields and remote sensing for forestry applications. Here, we show neural radiance fields offer a wide range of possibilities to improve upon existing remote sensing methods in forest monitoring. We present experiments that demonstrate their potential to: (1) express fine features of forest 3D structure, (2) fuse available remote sensing modalities and (3), improve upon 3D structure derived forest metrics. Altogether, these properties make neural fields an attractive computational tool with great potential to further advance the scalability and accuracy of forest monitoring programs.

Juan Castorena

In this work, we propose the use of a causal collider structured model to describe the underlying data generative process assumptions in disentangled representation learning. This extends the conventional i.i.d. factorization assumption model $p(\mathbf{y}) = \prod_{i} p(\mathbf{y}_i )$, inadequate to handle learning from biased datasets (e.g., with sampling selection bias). The collider structure, explains that conditional dependencies between the underlying generating variables may be exist, even when these are in reality unrelated, complicating disentanglement. Under the rubric of causal inference, we show this issue can be reconciled under the condition of causal identification; attainable from data and a combination of constraints, aimed at controlling the dependencies characteristic of the \textit{collider} model. For this, we propose regularization by identification (ReI), a modular regularization engine designed to align the behavior of large scale generative models with the disentanglement constraints imposed by causal identification. Empirical evidence on standard benchmarks demonstrates the superiority of ReI in learning disentangled representations in a variational framework. In a real-world dataset we additionally show that our framework, results in interpretable representations robust to out-of-distribution examples and that align with the true expected effect from domain knowledge.

Juan Castorena, Diane Oyen

In this work we propose a deep learning approach to clean spectroscopy signals using only uncleaned data. Cleaning signals from spectroscopy instrument noise is challenging as noise exhibits an unknown, non-zero mean, multivariate distributions. Our framework is a siamese neural net that learns identifiable disentanglement of the signal and noise components under a stationarity assumption. The disentangled representations satisfy reconstruction fidelity, reduce consistencies with measurements of unrelated targets and imposes relaxed-orthogonality constraints between the signal and noise representations. Evaluations on a laser induced breakdown spectroscopy (LIBS) dataset from the ChemCam instrument onboard the Martian Curiosity rover show a superior performance in cleaning LIBS measurements compared to the standard feature engineered approaches being used by the ChemCam team.

Juan Castorena, Diane Oyen, Ann Ollila et al.

This work proposes a spectral convolutional neural network (CNN) operating on laser induced breakdown spectroscopy (LIBS) signals to learn to (1) disentangle spectral signals from the sources of sensor uncertainty (i.e., pre-process) and (2) get qualitative and quantitative measures of chemical content of a sample given a spectral signal (i.e., calibrate). Once the spectral CNN is trained, it can accomplish either task through a single feed-forward pass, with real-time benefits and without any additional side information requirements including dark current, system response, temperature and detector-to-target range. Our experiments demonstrate that the proposed method outperforms the existing approaches used by the Mars Science Lab for pre-processing and calibration for remote sensing observations from the Mars rover, 'Curiosity'.

Manish Bhattarai, Diane Oyen, Juan Castorena et al.

Resolution of the complex problem of image retrieval for diagram images has yet to be reached. Deep learning methods continue to excel in the fields of object detection and image classification applied to natural imagery. However, the application of such methodologies applied to binary imagery remains limited due to lack of crucial features such as textures,color and intensity information. This paper presents a deep learning based method for image-based search for binary patent images by taking advantage of existing large natural image repositories for image search and sketch-based methods (Sketches are not identical to diagrams, but they do share some characteristics; for example, both imagery types are gray scale (binary), composed of contours, and are lacking in texture). We begin by using deep learning to generate sketches from natural images for image retrieval and then train a second deep learning model on the sketches. We then use our small set of manually labeled patent diagram images via transfer learning to adapt the image search from sketches of natural images to diagrams. Our experiment results show the effectiveness of deep learning with transfer learning for detecting near-identical copies in patent images and querying similar images based on content.

Juan Castorena, Manish Bhattarai, Diane Oyen

Binary image based classification and retrieval of documents of an intellectual nature is a very challenging problem. Variations in the binary image generation mechanisms which are subject to the document artisan designer including drawing style, view-point, inclusion of multiple image components are plausible causes for increasing the complexity of the problem. In this work, we propose a method suitable to binary images which bridges some of the successes of deep learning (DL) to alleviate the problems introduced by the aforementioned variations. The method consists on extracting the shape of interest from the binary image and applying a non-Euclidean geometric neural-net architecture to learn the local and global spatial relationships of the shape. Empirical results show that our method is in some sense invariant to the image generation mechanism variations and achieves results outperforming existing methods in a patent image dataset benchmark.

Juan Castorena, Gint Puskorius, Gaurav Pandey

This work proposes a novel motion guided method for target-less self-calibration of a LiDAR and camera and use the re-projection of LiDAR points onto the image reference frame for real-time depth upsampling. The calibration parameters are estimated by optimizing an objective function that penalizes distances between 2D and re-projected 3D motion vectors obtained from time-synchronized image and point cloud sequences. For upsampling, a simple, yet effective and time efficient formulation that minimizes depth gradients subject to an equality constraint involving the LiDAR measurements is proposed. Validation is performed on recorded real data from urban environments and demonstrations that our two methods are effective and suitable to mobile robotics and autonomous vehicle applications imposing real-time requirements is shown.

Juan Castorena, Siddharth Agarwal

In this work we propose an alternative formulation to the problem of ground reflectivity grid based localization involving laser scanned data from multiple LIDARs mounted on autonomous vehicles. The driving idea of our localization formulation is an alternative edge reflectivity grid representation which is invariant to laser source, angle of incidence, range and robot surveying motion. Such property eliminates the need of the post-factory reflectivity calibration whose time requirements are infeasible in mass produced robots/vehicles. Our experiments demonstrate that we can achieve better performance than state of the art on ground reflectivity inference-map based localization at no additional computational burden.

Juan Castorena

In this investigation we focus on the problem of mapping the ground reflectivity with multiple laser scanners mounted on mobile robots/vehicles. The problem originates because regions of the ground become populated with a varying number of reflectivity measurements whose value depends on the observer and its corresponding perspective. Here, we propose a novel automatic, data-driven computational mapping framework specifically aimed at preserving edge sharpness in the map reconstruction process and that considers the sources of measurement variation. Our new formulation generates map-perspective gradients and applies sub-set selection fusion and de-noising operators to these through iterative algorithms that minimize an $\ell_1$ sparse regularized least squares formulation. Reconstruction of the ground reflectivity is then carried out based on Poisson's formulation posed as an $\ell_2$ term promoting consistency with the fused gradient of map-perspectives and a term that ensures equality constraints with reference measurement data. We demonstrate our new framework outperforms the capabilities of existing ones with experiments realized on Ford's fleet of autonomous vehicles. For example, we show we can achieve map enhancement (i.e., contrast enhancement), artifact removal, de-noising and map-stitching without requiring an additional reflectivity adjustment to calibrate sensors to the specific mounting and robot/vehicle motion.