Razvan Stefanescu

Ahmed Attia, Razvan Stefanescu, Adrian Sandu

Hybrid Monte-Carlo (HMC) sampling smoother is a fully non-Gaussian four-dimensional data assimilation algorithm that works by directly sampling the posterior distribution formulated in the Bayesian framework. The smoother in its original formulation is computationally expensive due to the intrinsic requirement of running the forward and adjoint models repeatedly. Here we present computationally efficient versions of the HMC sampling smoother based on reduced-order approximations of the underlying model dynamics. The schemes developed herein are tested numerically using the shallow-water equations model on Cartesian coordinates. The results reveal that the reduced-order versions of the smoother are capable of accurately capturing the posterior probability density, while being significantly faster than the original full order formulation.

Azam Moosavi, Razvan Stefanescu, Adrian Sandu

This paper introduces multivariate input-output models to predict the errors and bases dimensions of local parametric Proper Orthogonal Decomposition reduced-order models. We refer to these multivariate mappings as the MP-LROM models. We employ Gaussian Processes and Artificial Neural Networks to construct approximations of these multivariate mappings. Numerical results with a viscous Burgers model illustrate the performance and potential of the machine learning based regression MP-LROM models to approximate the characteristics of parametric local reduced-order models. The predicted reduced-order models errors are compared against the multi-fidelity correction and reduced order model error surrogates methods predictions, whereas the predicted reduced-order dimensions are tested against the standard method based on the spectrum of snapshots matrix. Since the MP-LROM models incorporate more features and elements to construct the probabilistic mappings they achieve more accurate results. However, for high-dimensional parametric spaces, the MP-LROM models might suffer from the curse of dimensionality. Scalability challenges of MP-LROM models and the feasible ways of addressing them are also discussed in this study.

Razvan Stefanescu, Ethan Oh, Ruben Vazquez et al.

We introduce a multi-modal WAVE-DETR drone detector combining visible RGB and acoustic signals for robust real-life UAV object detection. Our approach fuses visual and acoustic features in a unified object detector model relying on the Deformable DETR and Wav2Vec2 architectures, achieving strong performance under challenging environmental conditions. Our work leverage the existing Drone-vs-Bird dataset and the newly generated ARDrone dataset containing more than 7,500 synchronized images and audio segments. We show how the acoustic information is used to improve the performance of the Deformable DETR object detector on the real ARDrone dataset. We developed, trained and tested four different fusion configurations based on a gated mechanism, linear layer, MLP and cross attention. The Wav2Vec2 acoustic embeddings are fused with the multi resolution feature mappings of the Deformable DETR and enhance the object detection performance over all drones dimensions. The best performer is the gated fusion approach, which improves the mAP of the Deformable DETR object detector on our in-distribution and out-of-distribution ARDrone datasets by 11.1% to 15.3% for small drones across all IoU thresholds between 0.5 and 0.9. The mAP scores for medium and large drones are also enhanced, with overall gains across all drone sizes ranging from 3.27% to 5.84%.

Vedant Bhasin, Matthew Yudin, Razvan Stefanescu et al.

Trojan backdoors can be injected into large language models at various stages, including pretraining, fine-tuning, and in-context learning, posing a significant threat to the model's alignment. Due to the nature of causal language modeling, detecting these triggers is challenging given the vast search space. In this study, we propose a multistage framework for detecting Trojan triggers in large language models consisting of token filtration, trigger identification, and trigger verification. We discuss existing trigger identification methods and propose two variants of a black-box trigger inversion method that rely on output logits, utilizing beam search and greedy decoding respectively. We show that the verification stage is critical in the process and propose semantic-preserving prompts and special perturbations to differentiate between actual Trojan triggers and other adversarial strings that display similar characteristics. The evaluation of our approach on the TrojAI and RLHF poisoned model datasets demonstrates promising results.

Todd Huster, Peter Lin, Razvan Stefanescu et al.

Neural networks can conceal malicious Trojan backdoors that allow a trigger to covertly change the model behavior. Detecting signs of these backdoors, particularly without access to any triggered data, is the subject of ongoing research and open challenges. In one common formulation of the problem, we are given a set of clean and poisoned models and need to predict whether a given test model is clean or poisoned. In this paper, we introduce a detector that works remarkably well across many of the existing datasets and domains. It is obtained by training a binary classifier on a large number of models' weights after performing a few different pre-processing steps including feature selection and standardization, reference model weights subtraction, and model alignment prior to detection. We evaluate this algorithm on a diverse set of Trojan detection benchmarks and domains and examine the cases where the approach is most and least effective.

Azam Moosavi, Razvan Stefanescu, Adrian Sandu

Reduced order models are computationally inexpensive approximations that capture the important dynamical characteristics of large, high-fidelity computer models of physical systems. This paper applies machine learning techniques to improve the design of parametric reduced order models. Specifically, machine learning is used to develop feasible regions in the parameter space where the admissible target accuracy is achieved with a predefined reduced order basis, to construct parametric maps, to chose the best two already existing bases for a new parameter configuration from accuracy point of view and to pre-select the optimal dimension of the reduced basis such as to meet the desired accuracy. By combining available information using bases concatenation and interpolation as well as high-fidelity solutions interpolation we are able to build accurate reduced order models associated with new parameter settings. Promising numerical results with a viscous Burgers model illustrate the potential of machine learning approaches to help design better reduced order models.