Engineering topics

Godfrey, Charles W.

Publications and source records attributed to Godfrey, Charles W..

4 records

DOE OSTI2023

Understanding the Inner-Workings of Language Models Through Representation Dissimilarity

We use model stitching to understand the internal representations of language models. Similar to vision models, we find that "more is better," and representations learned with more data and larger width can improve the performance of weaker models via stitching. We likewise find that certain architecture choices, using GeLU vs SoLU activation functions, influence the quality of learned representations. Finally, model stitching (as opposed to other model diagnostic methods, like mode connectivity) can localize the different generalization strategies of text classifiers under domain shift to certain hidden layers.

Brown, Davis R.↗

DOE OSTI2023

Neural Image Compression: Generalization, Robustness, and Spectral Biases

Recent advances in neural image compression (NIC) have resulted in models which are starting to outperform traditional codecs. While this has led to growing excitement about using these methods in real-world applications, the successful adoption of any machine learning system (including NIC) in the wild requires it to generalize (and be robust) to unseen distribution shifts at deployment time. Unfortunately, current research lacks comprehensive datasets and informative tools to evaluate and understand compression performance in real-world settings. To bridge this crucial gap, first, this paper presents a comprehensive benchmark suite to evaluate the out-of-distribution (OOD) performance of image compression methods. Specifically, we design CLIC-C and Kodak-C by introducing 15 common corruptions to popular CLIC and Kodak benchmarks. Next, we propose spectrally inspired introspection tools to gain a deeper understanding of errors introduced by image compression methods as well as their OOD performance. To this end, we carry out a detailed performance comparison of the classical codec with various variants of NIC (e.g., original, variable rate, pruned), revealing intriguing findings that challenge our current understanding of the strengths and limitations of NIC. Finally, we corroborate our empirical findings with theoretical analysis, providing an in-depth view of the OOD performance of NIC. Our benchmarks, spectral introspection tools, and findings provide a crucial bridge to the real-world adoption of NIC. We hope that our work will propel future efforts in designing more robust and generalizable NIC methods.

neural networks, Variational Autoencoder, robustne↗

DOE OSTI2023

Gumby: Quantifying multi-modal model resiliency

With the rise of cheap data and sensors, more use cases are emerging for multi-input models. Research has shown that including multiple data modalities can improve performance, suggesting that deep learning models can successfully learn to leverage complementary information from different modalities. However, this improved predictive power comes with unanticipated costs: additional inputs change model resiliency and expand the threat space for adversarial attacks. We first provide theoretical underpinnings for how adversarial success scales with input dimension. We then characterize the performance of a suite of multispectral deep learning models with different fusion approaches, quantify their relative reliance on different input bands, and evaluate their robustness to naturalistic and adversarial image corruptions.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2023

Quantifying the robustness of deep multispectral segmentation models against natural perturbations and data poisoning

In overhead image segmentation tasks, including additional spectral bands beyond the traditional RGB channels can improve model performance. However, it is still unclear how incorporating this additional data impacts model robustness to adversarial attacks and natural perturbations. For adversarial robustness, the additional in-formation could improve the model’s ability to distinguish malicious inputs, or simply provide new attack avenues and vulnerabilities. For natural perturbations, the additional information could better inform model decisions and weaken perturbation effects or have no significant influence at all. In this work, we seek to characterize the performance and robustness of a multispectral (RGB and near infrared) image segmentation model subjected to adversarial attacks and natural perturbations. While existing adversarial and natural robustness research has focused primarily on digital perturbations, we prioritize on creating realistic perturbations designed with physical world conditions in mind. For adversarial robustness, we focus on data poisoning attacks whereas for natural robustness, we focus on extending ImageNet-C common corruptions for fog and snow that coherently and self-consistently perturbs the input data. Overall, we find both RGB and multispectral models are vulnerable to data poisoning attacks regardless of input or fusion architectures and that while physically-realizable natural perturbations still degrade model performance, the impact differs based on fusion architecture and input data.

Deep learning, multispectral images, multimodal fu↗