Engineering Papers⌕ Search

Engineering topics

Divya Gopinath

Publications and source records attributed to Divya Gopinath.

17 records
All sourcesNASA NTRSDOE OSTI
NASA NTRS2024

Recommendations on Evidence and Process for Certification of Learning-enabled Components in Aerospace Systems

This report primarily identifies a collection of relevant and necessary evidence for assurance of machine learnt components (MLCs)—also known as learning-enabled components—integrated into aircraft systems, and gives preliminary suggestions on the elements of a certification process that invoke the identified evidence. The main focus is on feedforward neural networks that are static and trained offline through supervised learning. A brief background on the generic elements of the lifecycle of an MLC is given to contextualize the assurance considerations and, consequently, the evidence that is relevant and necessary to support certification. At the level of an MLC, those considerations relate to: (i) the consistency and correctness of MLC contributions to system functions in the context of a validated functional intent; and (ii) the absence of MLC contributions to aircraft-level failure conditions. At an ML model level, confidence in model and data properties contribute to assurance of the containing MLC, in particular: (a) generalizability and robustness of models, in the presence of inputs not previously seen during training, disturbances to inputs, and unexpected inputs; and (b) valid data, i.e., data that are at least representative, relevant, complete, and accurate. Evidence for the above span the elements of the ML lifecycle, and includes, at a minimum, lifecycle artifacts that pertain to: (1) properties of requirements capturing functional intent, safety constraints, and aspects of the intended use and operating environment; (2) model performance, model complexity and design, and algorithm choice; (3) achievement of required performance at the levels of a trained model during model development, a trained model after model development is complete, and a trained model that is transformed into an executable equivalent; (4) model implementation aspects necessary for transforming a trained model into the executable equivalent; (5) integration of the executable trained model into the containing MLC, and eventually the larger system; and, (6) lastly, the verification and validation (V&V) of each of the above. Such V&V lifecycle artifacts themselves include: aspects of coverage, e.g., of various levels of requirements by the input space of the model and the data; traceability (where applicable); application of formal methods for property specification, analysis, and checking. Examples of evidence generation methods and tools further ground the discussion on what constitutes evidence, and the contribution to assurance during certification. The identified assurance considerations and supporting evidence is not a comprehensive set. Additionally, neither what should be considered as sufficient evidence relative to the assigned criticality of an MLC, nor how criticality ought to be determined and adjusted, have been considered in this report. However, suggestions are made for potential activities of the ML lifecycle that are aimed at providing confidence that an MLC can be relied upon when integrated into its containing (aircraft) system. Those activities are proposed as candidate elements of a certification process for MLCs. The main purpose of this report to inform regulatory guidance and consensus standards that may be used to meet the safety intent of the applicable regulations.

Aviation safety↗
NASA NTRSDate not supplied

On the Probabilistic Analysis of Neural Networks

Neural networks are powerful tools for automated decision-making,seeing increased application in safety-critical domains, such as autonomous driving. Due to their black-box nature and large scale,reasoning about their behavior is challenging. Statistical analysis is often used to infer probabilistic properties of a network, such as its robustness to noise and inaccurate inputs. While scalable, statistical methods can only provide probabilistic guarantees on the quality of their results and may underestimate the impact of low probability inputs leading to undesired behavior of the network.We investigate here the use of symbolic analysis and constraint solution space quantification to precisely quantify probabilistic properties in neural networks. We demonstrate the potential of the proposed technique in a case study involving the analysis of ACAS-Xu, a collision avoidance system for unmanned aircraft control.

neural nets↗
NASA NTRSDate not supplied

NEUROSPF: A Tool For the Symbolic Analysis of Neural Networks

This paper presents NEUROSPF, a tool for the symbolic analysis of neural networks. Given a trained neural network model, the tool extracts the architecture and model parameters and translates them into a Java representation that is amenable for analysis using the Symbolic PathFinder symbolic execution tool. Notably, NEUROSPF encodes specialized peer classes for parsing the model’s parameters, thereby enabling efficient analysis. With NEUROSPF the user has the flexibility to specify either the inputs or the network internal parameters as symbolic, promoting the application of program analysis and testing approaches from software engineering to the field of machine learning. For instance, NEUROSPF can be used for coverage-based testing and test generation, finding adversarial examples and also constraint-based repair of neural networks, thus improving the reliability of neural networks and of the applications that use them.

neural networks↗
NASA NTRSDate not supplied

QuantifyML: How good is my machine learning model?

This paper presents an approach, QuantifyML, which employs model counting to assess the learnability and robustness of machine learning models. Typically the efficacy of machine learning models is determined by computing their accuracy statistically on test data sets. However, this may be misleading, if the test data is not representative of the problem that is being studied. Further, two different models may have the same accuracy on a given data set, measured statistically, but may be very different in their behavior on unseen data. Also, models with high accuracy could have poor adversarial robustness. In QuantifyML, our goal is to precisely quantify the extent to which machine learning models have learned and generalized from the given data. In QuantifyML, a trained model is translated into a C program, which is fed to the CBMC model checking tool to produce a formula in Conjunctive Normal Form (CNF), which in turn is analyzed with state-of-the-art model counters to efficiently obtain precise counts w.r.t different outputs. QuantifyML enables i) evaluating the learnability of models by comparing the counts for the outputs to ground truth, expressed as logical predicates (if available), ii) comparing the performance of different models that may be built with different machine learning algorithms (e.g., decision-trees vs. neural networks), and iii) quantifying the robustness of trained models around given inputs. Our evaluation demonstrates these applications of QuantifyML on decision trees and neural networks trained to learn relational properties of graphs, for which we know the ground truth, and to perform image classification, for which we do not have the ground truth, but we can quantify local robustness.

Deep Neural Networks↗
NASA NTRSDate not supplied

QuantifyML: How Good is My Machine Learning Model?

We present QuantifyML, which applies model counting to assess the learn ability, safety, and robustness of machine learning models. Typically, the efficacy of machine learning models is determined by computing their accuracy statistically on test datasets. However, this may be misleading, if the test data is not representative of the problem that is being studied. With QuantifyML we aim to precisely quantify the extent to which machine learning models have learned and generalized from the given data. In QuantifyML, a trained model is translated into aC program, which is fed to the CBMC model checking tool to produce a formula in Conjunctive Normal Form (CNF), which in turn is analyzed with state-of-the-art model counters to efficiently obtain precise countsw.r.t different outputs. QuantifyML enables i) evaluating the learnability of models by comparing the counts for the outputs to ground truth, ex-pressed as logical predicates (if available), ii) comparing the performance of different models that may be built with different machine learning algorithms (e.g., decision-trees vs. neural networks), and iii) quantifying the safety and robustness of trained models.

Deep Neural Networks↗
NASA NTRSDate not supplied

Case Study: Analysis of Autonomous Center line Tracking Neural Networks

Deep neural networks have gained widespread usage in a number of applications. However, limitations such as lack of explainability and robustness inhibit building trust in their behavior, which is crucial in safety critical applications such as autonomous driving. Therefore, techniques which aid in understanding and providing guarantees for neural network behavior are the need of the hour. In this paper, we present a case study applying a recently proposed technique, Prophecy, to analyze the behavior of a neural network model, provided by our industry partner and used for autonomous guiding of airplanes on taxi runways. This regression model takes as input an image of the runway and produces two outputs, cross-track error and heading error, which represent the position of the plane relative to the center line. We use the Prophecy tool to extract neuron activation patterns for the correctness and safety properties of the model. We show the use of these patterns to identify features of the input that explain correct and incorrect behavior. We also use the patterns to provide guarantees of consistent behavior. We explore a novel idea of using sequences of images (instead of single images) to obtain good explanations and identify regions of consistent behavior.

Deep Neural Networks↗
NASA NTRSDate not supplied

NNREPAIR: Constraint-based Repair of Neural Network Classifiers

We present NNREPAIR, a constraint-based technique for repairing neural network classifiers by i) fixing the logic of the network at an intermediate layer, and ii) fixing the decision conditions of the network at the last layer. NNREPAIR first uses fault localization to find potentially faulty network parameters (specifically weights) and then performs repair using constraint solving to apply small modifications to remedy the defects. We present novel strategies to enable precise yet efficient repair such as inferring correctness specifications to act as oracles for intermediate layer repair, and generation of experts wrt each class. We demonstrate the technique in the context of three different scenarios: (1) Improving the over all accuracy of a model, (2) Fixing security vulnerabilities caused by poisoning of training data and (3) Improving the robustness of the network against adversarial attacks. Our evaluation on MNIST and CIFAR-10 models shows that NNREPAIR can improve the accuracy by 45.56% on poisoned data and 11.11% on adversarial data. NNREPAIR also provides improvement, albeit modest, in the overall accuracy of models, without requiring new data or re-training.

Repair↗
NASA NTRSDate not supplied

Understanding and Verifying Neural Networks

Deep Neural Networks (DNNs) have gained immense popularity in recent times and have widespread use in applications such as image classification, sentiment analysis, speech recognition and also in safety-critical applications such as autonomous driving. However, they suffer limitations such as lack of explainability and robustness which raise safety and security concerns in their usage. Further, the complex structure and large input spaces of DNNs act as an impediment to thorough verification and testing. The SafeDNN project at the Robust Software Engineering (RSE) group at NASA aims at exploring techniques to ensure that systems that use deep neural networks are safe, robust and interpretable. In this talk, I will be presenting our technique Prophecy that automatically infers formal properties of deep neural network models. The tool extracts patterns based on neuron activations as preconditions that imply certain desirable output properties of the model. I would be highlighting case studies that use Prophecy in obtaining explanations for network decisions, understanding correct and incorrect behavior, providing formal guarantees wrt safety and robustness, and debugging neural network models. We have applied the tool on image classification networks, neural network controllers providing turn advisories in unmanned aircrafts, regression models used for autonomous center-line tracking in aircrafts and neural network object detectors

Deep Neural Networks↗
NASA NTRSDate not supplied

Feature-Guided Analysis of Neural Networks

Applying standard software engineering practices to neural networks is challenging due to the lack of high-level abstractions describing a neural network’s behavior. To address this challenge, we propose to extract high-level task-specific features from the neural network internal representation, based on monitoring the neural network activations.The extracted feature representations can serve as a link to high-level requirements and can be leveraged to enable fundamental software engineering activities, such as automated testing, debugging, requirements analysis, and formal verification, leading to better engineering of neural networks. Using two case studies, we present initial empirical evidence demonstrating the feasibility of our ideas.

Features↗
NASA NTRSDate not supplied

Assumption Generation for the Verification of Learning-Enabled Autonomous Systems

Providing safety guarantees for autonomous systems is difficultas these systems operate in complex environments that require the use of learning-enabled components, such as deep neural networks (DNNs) for visual perception. DNNs are hard to analyze due to their size (they can have thousands or millions of parameters), lack of formal specifications (DNNs are typically learnt from labeled data, in the absence of any formal or informal requirements), and sensitivity to small changes in the environment. We present an assume-guarantee style compositional approach for the formal verification of system-level safety properties of such autonomous systems. Our insight is that we can analyze the system in the absence of the DNN perception components by automatically synthesizing assumptions on the DNN behaviour that guarantee the satisfaction of the required safety properties. The synthesized assumptions are the weakest in the sense that they characterize the output sequences of all the possible DNNs that, plugged into the autonomous system, guarantee the required safety properties. The assumptions can be leveraged as run-time monitors over a deployed DNN to guarantee the safety of the overall system; they can also be mined to extract local specifications for use during training and testing of DNNs. We illustrate our approach on a case study taken from the autonomous airplanes domain that uses a complex DNN for perception

Autonomous systems↗
NASA NTRSDate not supplied

Evaluating the Trustworthiness of Deep Neural Networks in Deployment – A Comparative Study (Replicability Study)

As deep neural networks (DNNs) are increasingly used in safety critical applications, there is a growing concern for their trustworthiness. Even highly trained, high-performant networks are not 100% accurate. However, it is very difficult to predict their behaviour during deployment without ground truth. In this paper, we provide a comparative and replicability study on recent approaches that have been proposed to evaluate the trustworthiness of DNNs. We find that it is very difficult to run and reproduce the results for these approaches on their replication packages, and it is even more difficult to run the tools on artifacts other than their own. Further, it is difficult to compare the effectiveness of the tools, due to lack of clearly defined evaluation metrics. Our results indicate that more effort is needed in our research community to obtain sound techniques for evaluating the trustworthiness of neural networks in safety-critical domains. To this end, we contribute an evaluation framework that incorporates the considered approaches and enables evaluation on common benchmarks, using common metrics. Using this framework, we run a comparative study of the three approaches.

Trustworthy AI↗
NASA NTRSDate not supplied

Concept-based Analysis of Neural Networks via Vision-Language Models

The analysis of vision-based deep neural networks (DNNs) is highly desirable but it is very challenging due to the difficulty of expressing formal specifications for vision tasks and the lack of efficient verification procedures. In this paper, we propose to leverage emerging multimodal, vision-language, foundation models (VLMs) as a lens through which we can reason about vision models. VLMs have been trained on a large body of images accompanied by their textual description, and are thus implicitly aware of high-level, human-understandable concepts describing the images. We describe a logical specification language Con spec designed to facilitate writing specifications in terms of these concepts. To define and formally check Con spec specifications, we build a map between the internal representations of a given vision model and a VLM, leading to an efficient verification procedure of natural-language properties for vision models. We demonstrate our techniques on a ResNet-based classifier trained on the RIVAL-10 dataset using CLIP as the multimodal model.

Large Vision Language Models↗