Results for “Provable repair”

DOE OSTI2024

Provable Repair of Vision Transformers

Vision Transformers have emerged as state-of-the-art image recognition tools, but may still exhibit incorrect behavior. Incorrect image recognition can have disastrous consequences in safety-critical real-world applications such as self-driving automobiles. In this paper, we present Provable Repair of Vision Transformers (PRoViT), a provable repair approach that guarantees the correct classification of images in a repair set for a given Vision Transformer without modifying its architecture. PRoViT avoids negatively affecting correctly classified images (drawdown) by minimizing the changes made to the Vision Transformer’s parameters and original output. Here, we observe that for Vision Transformers, unlike for other architectures such as ResNet or VGG, editing just the parameters in the last layer achieves correctness guarantees and very low drawdown. We introduce a novel method for editing these last-layer parameters that enables PRoViT to efficiently repair state-of-the-art Vision Transformers for thousands of images, far exceeding the capabilities of prior provable repair approaches.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2023

Architecture-Preserving Provable Repair of Deep Neural Networks

Deep neural networks (DNNs) are becoming increasingly important components of software, and are considered the state-of-the-art solution for a number of problems, such as image recognition. However, DNNs are far from infallible, and incorrect behavior of DNNs can have disastrous real-world consequences. This paper addresses the problem of architecture-preserving V-polytope provable repair of DNNs. A V-polytope defines a convex bounded polytope using its vertex representation. V-polytope provable repair guarantees that the repaired DNN satisfies the given specification on the infinite set of points in the given V-polytope. An architecture-preserving repair only modifies the parameters of the DNN, without modifying its architecture. The repair has the flexibility to modify multiple layers of the DNN, and runs in polynomial time. It supports DNNs with activation functions that have some linear pieces, as well as fully-connected, convolutional, pooling and residual layers. To the best our knowledge, this is the first provable repair approach that has all of these features. We implement our approach in a tool called APRNN. Using MNIST, ImageNet, and ACAS Xu DNNs, we show that it has better efficiency, scalability, and generalization compared to PRDNN and REASSURE, prior provable repair methods that are not architecture preserving.

97 MATHEMATICS AND COMPUTING↗

NASA NTRSDate not supplied

Predicting SLS Launch Environment using a Novel Multiphase Formulation

Powerful acoustic waves generated during ignition of launch vehicles may be dangerous to the vehicle, its payload, or the surrounding structures. The water-based Ignition Overpressure and Sound Suppression (IOP/SS) system at Kennedy Space Center’s (KSC) Launch Complex 39B (LC-39B) will be used to protect the Space Launch System (SLS) from the acoustic vibrations generated during launch. The IOP/SS system uses enormous amounts of water to dampen and attenuate these sound waves. To better understand the launch environment risks and to study the effectiveness of the IOP/SS system it is desirable to have time-accurate unsteady simulations of the vehicle ignition with water-based sound suppression. This paper presents results obtained with a novel, high-order accurate, and robust numerical method designed for simulating compressible multiphase flows. A positivity-preserving finite difference scheme is utilized which is formally high-order accurate and also provably robust. Robustness is critical due to the extreme nature of the flow which exhibits highly nonlinear shock and rarefaction waves interacting with liquid-gas interfaces with density ratios of the order of 1000:1. Furthermore, the high-order accuracy (and the high resolution property) is desirable for predicting wave phenomena like IOP waves since the signal can be resolved accurately and propagated long distances with fewer grid points. This finite-difference method was developed using NASA’s Launch, Ascent, and Vehicle Aerodynamics (LAVA) Cartesian immersed boundary framework. We present a validation case by applying our solver to the SLS Scale Model Acoustic Test (SMAT). The SLS SMAT is a well-instrumented 5% scale model test meant to represent the SLS at NASA KSC’s LC-39B pad. Scale IOP tests were performed with and without the sound suppression water and included many sensors which recorded the pressure waves produced during ignition. For this validation case we conduct two simulations, likewise with and without sound suppression water, and compare the SLS SMAT pressure sensor signals with our numerical signals at identical locations. Following this validation case we present a study of the SLS launch environment to examine engineering safety concerns about the mobile launch pad. Engineers at KSC redesigned the main flame deflector at LC-39B anticipating the increased loads from the SLS and to repair damage from prior Shuttle missions. This deflector redesign made use of surface pressure and temperature data from LAVA full-scale SLS simulations without the sound suppression system. The engineers were questioning the possibility of increased pressure loads on the underside of the mobile launcher due to the water in the flame trench. Based on the results established in our simulations of the SLS SMAT, we performed updated calculations for SLS at LC-39B with and without water systems active to assess the readiness of the launch pad for Artemis I launch. Our results show that the IOP/SS system is effective at reducing the overpressure signal and overall sound pressure levels felt by the vehicle and additionally that the pressure loads experienced by the mobile launcher (ML) during engine startup is not increased by the presence of water.

EGS↗