Paul Gavrikov

I am a PhD student at the Institute of Machine Learning Analytics, part of the Offenburg University, Germany where I work on fundamental understanding of Computer Vision models. I am supervised by Janis Keuper.

Email  /  X (Twitter)  /  GitHub  /  Google Scholar  /  LinkedIn  /  Medium  /  dblp

News

Research

My research is bridging mechanistic interpretability with training dynamics, robustness, and generalization. My published works extensively explored learned weights, such as filters in CNNs, and inherent biases, like texture/shape or frequency. Recently, I shifted to investigating (vision-language) foundation models. A long time ago (but not in a galaxy far, far away), I worked on Bluetooth Low Energy. Highlights.

Are Vision Language Models Texture or Shape Biased and Can We Steer Them?

Surprisingly, LLM-powered Vision-Language Models (VLMs), often exhibit more shape bias than pure vision models, influenced by language. We show that this allows steering a purely visual bias by language.

Can Biases in ImageNet Models Explain Generalization?

We investigate the generalization capabilities of neural networks from the perspective of shape bias, spectral biases, and the critical band. Our results show that even when we fix the architecture these indicators are not reliable predictors of generalization performance.

Improving Native CNN Robustness with Filter Frequency Regularization

We propose controlling the frequency content of learned convolution filters in vision CNNs. This results in model that are natively more robust to adversarial robustness and corruptions, generalize better, and are generally more aligned with human vision.

Don't Look into the Sun: Adversarial Solarization Attacks on Image Classifiers

We present a new adversarial attack based on image solarization. Despite being conceptually simple, the attack is effective, cheap to compute, and does not risk destroying the global structure of natural images. It also serves as a universal black-box attack against models trained with the legacy ImageNet training recipe.

On the Interplay of Convolutional Padding and Adversarial Robustness

Our study examines the relationship between padding in Convolutional Neural Networks (CNNs) and vulnerabilities to adversarial attacks. We show that adversarial attacks result in different perturbation anomalies at image boundaries depending on the padding mode and discuss which mode is the best for adversarial settings. (Spoiler: it's zero padding)

An Extended Study of Human-like Behavior under Adversarial Training

Our analysis reveals how different forms of adversarial training (AT) affect human-like behavior of CNNs and Transformers. Additionally, we propose a hypothesis of why AT increases shape bias and in which scenarios it can improve out-of-distribution generalization from a frequency perspective.

The Power of Linear Combinations: Learning with Random Convolutions

We question if learning spatial convolution filters is necessary. Even with default i.i.d. random inits, we can achieve 75.66% validation acc with ResNet-50 on ImageNet without ever learning any spatial convolution weight. Additionally, random filters can be more robust against adversarial attacks than learned filters.

Does Medical Imaging learn different Convolution Filters?

Earlier, we showed that, on average, convolution filters only show minor drifts when comparing various dimensions, including the learned task, image domain, or dataset. However, medical imaging models showed significant outliers through spiky filter distributions. We revisit this observation and perform an in-depth analysis of medical imaging models.

Robust Models are less Over-Confident

Adversarial Training (AT) leads to models that are significantly less overconfident with their decisions, even on clean data, than non-robust models. The analysis shows that not only AT, but also the models' building blocks (like activation functions and pooling) have a strong influence on the models' prediction confidences.

Adversarial Robustness through the Lens of Convolutional Filters

We investigate 3x3 convolution filters that form in adversarially-trained robust models and find that these models form more diverse, less sparse, and more orthogonal filters than their normal counterparts. The largest differences are found in the deepest layers and the very first convolution layer which forms highly distinct thresholding filters.

CNN Filter DB: An Empirical Investigation of Trained Convolutional Filters

We collected and publicly provided a dataset with over 1.4 billion 3x3 convolution filters from hundreds of trained CNNs. Our observations show that - surprisingly - models learn highly similar filter pattern distributions independent of task and dataset, but differ by model architecture. We also propose methods to measure the quality of filters to detect overparameterization or underfitting and show that many publicly available models suffer from "degenerated" filters.

An Empirical Investigation of Model-to-Model Distribution Shifts in Trained Convolutional Filters

This paper looks at distribution shifts in filter weights used for various computer vision tasks. We collected data from hundreds of trained CNNs and analyzed the distribution shifts along different axes of meta-parameters. We found interesting distribution shifts between trained filters, and argue that this is a valuable source for further investigation into understanding the impact of shifts in the input data on the generalization abilities of CNN models.

A Low Power and Low Latency Scan Algorithm for Bluetooth Low Energy Radios with Energy Detection Mechanisms

A new, more efficient algorithm for Bluetooth scanning is presented that uses less power and can scale with incoming network traffic. The algorithm does not require any changes to advertisers, so it is compatible with existing devices, and performance evaluation shows that it is more efficient than existing methods.

Exploring non-idealities in real device implementations of Bluetooth Mesh

We compare the performance of Bluetooth Mesh implementations on real chipsets against the ideal implementation of the specification. It is shown that there are non-idealities in the underlying Bluetooth Low Energy specification in real chipsets and in the implementation of Mesh, which introduces an unruly transmission as well as reception behavior. These effects lead to an impact on transmission rate, reception rate, latency, as well as a more significant impact on the average power consumption.

Using Bluetooth Low Energy to trigger a robust ultra-low power FSK wake-up receiver

We discuss a new approach to using BLE packets to create an FSK-like addressable wake-up packet. A wake-up receiver system was developed from off-the-shelf components to detect these packets. This system is more robust than traditional OOK wake-up systems and has a sensitivity of -47.8 dBm at a power consumption of 18.5 uW during passive listening. The system has a latency of 31.8 ms with a symbol rate of 1437 Baud.

Academic Career

Design and source code from Jon Barron's website