Advancements in Deep Neural Network Testing and Evaluation

The field of deep neural networks (DNNs) is rapidly evolving, with a significant focus on improving testing and evaluation methods. Recent developments have led to the creation of innovative approaches for test selection, input generation, and model assessment. Notably, researchers are exploring ways to adapt testing methods to fine-tuned models, which is crucial for ensuring the reliability of DNNs in real-world applications. Furthermore, the development of new metrics and techniques for evaluating DGMs has enhanced the ability to assess their performance and identify potential weaknesses. The use of latent space interpolation and representation improvement in latent space has shown promising results in generating realistic and diverse test inputs. Additionally, the application of large language models has enabled the automatic discovery of effective and diverse vulnerabilities in autonomous driving policies. Overall, these advancements are moving the field towards more robust and reliable DNNs. Noteworthy papers include: MetaSel, which introduces a novel test selection approach for fine-tuned DNN models, and PALATE, which proposes a holistic evaluation framework for deep generative models. HingeRLC-GAN is also notable for its ability to combat mode collapse in GANs, and AED for its framework that uses large language models to discover effective and diverse vulnerabilities in autonomous driving policies.

Sources

MetaSel: A Test Selection Approach for Fine-tuned DNN Models

Generating Realistic, Diverse, and Fault-Revealing Inputs with Latent Space Interpolation for Testing Deep Neural Networks

PALATE: Peculiar Application of the Law of Total Expectation to Enhance the Evaluation of Deep Generative Models

Latent Space Class Dispersion: Effective Test Data Quality Assessment for DNNs

HingeRLC-GAN: Combating Mode Collapse with Hinge Loss and RLC Regularization

Network Inversion for Generating Confidently Classified Counterfeits

Representation Improvement in Latent Space for Search-Based Testing of Autonomous Robotic Systems

AED: Automatic Discovery of Effective and Diverse Vulnerabilities for Autonomous Driving Policy with Large Language Models

Built with on top of