Not all learnable parameters (e.g., weights) contribute equally to a neural network's decision function. In fact, entire layers' parameters can sometimes be reset to random values with little to no impact on the model's decisions. We revisit earlier studies that examined how architecture and task complexity influence this phenomenon and ask: is this phenomenon also affected by how we train the model? We conducted experimental evaluations on a diverse set of ImageNet-1k classification models to explore this, keeping the architecture and training data constant but varying the training pipeline. Our findings reveal that the training method strongly influences which layers become critical to the decision function for a given task. For example, improved training regimes and self-supervised training increase the importance of early layers while significantly under-utilizing deeper layers. In contrast, methods such as adversarial training display an opposite trend. Our preliminary results extend previous findings, offering a more nuanced understanding of the inner mechanics of neural networks. Code: https://github.com/paulgavrikov/layer_criticality

The robust generalization of models to rare, in-distribution (ID) samples drawn from the long tail of the training distribution and to out-of-training-distribution (OOD) samples is one of the major challenges of current deep learning methods. For image classification, this manifests in the existence of adversarial attacks, the performance drops on distorted images, and a lack of generalization to concepts such as sketches. The current understanding of generalization in neural networks is very limited, but some biases that differentiate models from human vision have been identified and might be causing these limitations. Consequently, several attempts with varying success have been made to reduce these biases during training to improve generalization. We take a step back and sanity-check these attempts. Fixing the architecture to the well-established ResNet-50, we perform a large-scale study on 48 ImageNet models obtained via different training methods to understand how and if these biases - including shape bias, spectral biases, and critical bands - interact with generalization. Our extensive study results reveal that contrary to previous findings, these biases are insufficient to accurately predict the generalization of a model holistically. We provide access to all checkpoints and evaluation code at https://github.com/paulgavrikov/biases_vs_generalization

Vision language models (VLMs) have drastically changed the computer vision model landscape in only a few years, opening an exciting array of new applications from zero-shot image classification, over to image captioning, and visual question answering. Unlike pure vision models, they offer an intuitive way to access visual content through language prompting. The wide applicability of such models encourages us to ask whether they also align with human vision -- specifically, how far they adopt human-induced visual biases through multimodal fusion, or whether they simply inherit biases from pure vision models. One important visual bias is the texture vs. shape bias, or the dominance of local over global information. In this paper, we study this bias in a wide range of popular VLMs. Interestingly, we find that VLMs are often more shape-biased than their vision encoders, indicating that visual biases are modulated to some extent through text in multimodal models. If text does indeed influence visual biases, this suggests that we may be able to steer visual biases not just through visual input but also through language: a hypothesis that we confirm through extensive experiments. For instance, we are able to steer shape bias from as low as 49% to as high as 72% through prompting alone. For now, the strong human bias towards shape (96%) remains out of reach for all tested VLMs.

Assessing the robustness of deep neural networks against out-of-distribution inputs is crucial, especially in safety-critical domains like autonomous driving, but also in safety systems where malicious actors can digitally alter inputs to circumvent safety guards. However, designing effective out-of-distribution tests that encompass all possible scenarios while preserving accurate label information is a challenging task. Existing methodologies often entail a compromise between variety and constraint levels for attacks and sometimes even both. In a first step towards a more holistic robustness evaluation of image classification models, we introduce an attack method based on image solarization that is conceptually straightforward yet avoids jeopardizing the global structure of natural images independent of the intensity. Through comprehensive evaluations of multiple ImageNet models, we demonstrate the attack's capacity to degrade accuracy significantly, provided it is not integrated into the training augmentations. Interestingly, even then, no full immunity to accuracy deterioration is achieved. In other settings, the attack can often be simplified into a black-box attack with model-independent parameters. Defenses against other corruptions do not consistently extend to be effective against our specific attack. Project website: https://github.com/paulgavrikov/adversarial_solarization

It is common practice to apply padding prior to convolution operations to preserve the resolution of feature-maps in Convolutional Neural Networks (CNN). While many alternatives exist, this is often achieved by adding a border of zeros around the inputs. In this work, we show that adversarial attacks often result in perturbation anomalies at the image boundaries, which are the areas where padding is used. Consequently, we aim to provide an analysis of the interplay between padding and adversarial attacks and seek an answer to the question of how different padding modes (or their absence) affect adversarial robustness in various scenarios.

Following the traditional paradigm of convolutional neural networks (CNNs), modern CNNs manage to keep pace with more recent, for example transformer-based, models by not only increasing model depth and width but also the kernel size. This results in large amounts of learnable model parameters that need to be handled during training. While following the convolutional paradigm with the according spatial inductive bias, we question the significance of \emph{learned} convolution filters. In fact, our findings demonstrate that many contemporary CNN architectures can achieve high test accuracies without ever updating randomly initialized (spatial) convolution filters. Instead, simple linear combinations (implemented through efficient $1\times 1$ convolutions) suffice to effectively recombine even random filters into expressive network operators. Furthermore, these combinations of random filters can implicitly regularize the resulting operations, mitigating overfitting and enhancing overall performance and robustness. Conversely, retaining the ability to learn filter updates can impair network performance. Lastly, although we only observe relatively small gains from learning $3\times 3$ convolutions, the learning gains increase proportionally with kernel size, owing to the non-idealities of the independent and identically distributed (\textit{i.i.d.}) nature of default initialization techniques.

Neural networks have a number of shortcomings. Amongst the severest ones is the sensitivity to distribution shifts which allows models to be easily fooled into wrong predictions by small perturbations to inputs that are often imperceivable to humans and do not have to carry semantic meaning. Adversarial training poses a partial solution to address this issue by training models on worst-case perturbations. Yet, recent work has also pointed out that the reasoning in neural networks is different from humans. Humans identify objects by shape, while neural nets mainly employ texture cues. Exemplarily, a model trained on photographs will likely fail to generalize to datasets containing sketches. Interestingly, it was also shown that adversarial training seems to favorably increase the shift toward shape bias. In this work, we revisit this observation and provide an extensive analysis of this effect on various architectures, the common $\ell_2$- and $\ell_\infty$-training, and Transformer-based models. Further, we provide a possible explanation for this phenomenon from a frequency perspective.

We present first empirical results from our ongoing investigation of distribution shifts in image data used for various computer vision tasks. Instead of analyzing the original training and test data, we propose to study shifts in the learned weights of trained models. In this work, we focus on the properties of the distributions of dominantly used 3 3 convolution filter kernels. We collected and publicly provide a data set with over half a billion filters from hundreds of trained CNNs, using a wide range of data sets, architectures, and vision tasks. Our analysis shows interesting distribution shifts (or the lack thereof) between trained filters along different axes of meta-parameters, like data type, task, architecture, or layer depth. We argue, that the observed properties are a valuable source for further investigation into a better understanding of the impact of shifts in the input data to the generalization abilities of CNN models and novel methods for more robust transfer-learning in this domain.