Grigorios G. Chrysos

Thomas Pethick, Grigorios G. Chrysos, Volkan Cevher

Despite progress in adversarial training (AT), there is a substantial gap between the top-performing and worst-performing classes in many datasets. For example, on CIFAR10, the accuracies for the best and worst classes are 74% and 23%, respectively. We argue that this gap can be reduced by explicitly optimizing for the worst-performing class, resulting in a min-max-max optimization formulation. Our method, called class focused online learning (CFOL), includes high probability convergence guarantees for the worst class loss and can be easily integrated into existing training setups with minimal computational overhead. We demonstrate an improvement to 32% in the worst class accuracy on CIFAR10, and we observe consistent behavior across CIFAR100 and STL10. Our study highlights the importance of moving beyond average accuracy, which is particularly important in safety-critical applications.

Aggelina Chatziagapi, Grigorios G. Chrysos, Dimitris Samaras

We introduce MIGS (Multi-Identity Gaussian Splatting), a novel method that learns a single neural representation for multiple identities, using only monocular videos. Recent 3D Gaussian Splatting (3DGS) approaches for human avatars require per-identity optimization. However, learning a multi-identity representation presents advantages in robustly animating humans under arbitrary poses. We propose to construct a high-order tensor that combines all the learnable 3DGS parameters for all the training identities. By assuming a low-rank structure and factorizing the tensor, we model the complex rigid and non-rigid deformations of multiple subjects in a unified network, significantly reducing the total number of parameters. Our proposed approach leverages information from all the training identities and enables robust animation under challenging unseen poses, outperforming existing approaches. It can also be extended to learn unseen identities.

James Oldfield, Markos Georgopoulos, Grigorios G. Chrysos et al.

The Mixture of Experts (MoE) paradigm provides a powerful way to decompose dense layers into smaller, modular computations often more amenable to human interpretation, debugging, and editability. However, a major challenge lies in the computational cost of scaling the number of experts high enough to achieve fine-grained specialization. In this paper, we propose the Multilinear Mixture of Experts ($μ$MoE) layer to address this, focusing on vision models. $μ$MoE layers enable scalable expert specialization by performing an implicit computation on prohibitively large weight tensors entirely in factorized form. Consequently, $μ$MoEs (1) avoid the restrictively high inference-time costs of dense MoEs, yet (2) do not inherit the training issues of the popular sparse MoEs' discrete (non-differentiable) expert routing. We present both qualitative and quantitative evidence that scaling $μ$MoE layers when fine-tuning foundation models for vision tasks leads to more specialized experts at the class-level, further enabling manual bias correction in CelebA attribute classification. Finally, we show qualitative results demonstrating the expert specialism achieved when pre-training large GPT2 and MLP-Mixer models with parameter-matched $μ$MoE blocks at every layer, maintaining comparable accuracy. Our code is available at: https://github.com/james-oldfield/muMoE.

Jeffrey Wang, Jonathan Gregory, Grigorios G. Chrysos

Modern vision backbones treat pointwise activations (e.g., ReLU, GELU) and exponential softmax as essential sources of nonlinearity, but we demonstrate they are not required within MetaFormer-style vision backbones. We design activation-free polynomial alternatives for three core primitives (MLPs, convolutions, and attention), where Hadamard products replace standard nonlinearities to yield polynomial functions of the input. These modules integrate seamlessly into existing architectures: instantiated within MetaFormer, a modular framework for vision backbones, our PolyNeXt models match or exceed activation-based counterparts across model scales on ImageNet classification, ADE20K semantic segmentation, and out-of-distribution robustness. We also substantially outperform prior polynomial networks at reduced computational cost, showing that polynomial variants of standard modules beat complex custom architectures.

Leyla Naz Candogan, Yongtao Wu, Elias Abad Rocamora et al.

Defending aligned Large Language Models (LLMs) against jailbreaking attacks is a challenging problem, with existing approaches requiring multiple requests or even queries to auxiliary LLMs, making them computationally heavy. Instead, we focus on detecting jailbreaking input in a single forward pass. Our method, called Single Pass Detection SPD, leverages the information carried by the logits to predict whether the output sentence will be harmful. This allows us to defend in just one forward pass. SPD can not only detect attacks effectively on open-source models, but also minimizes the misclassification of harmless inputs. Furthermore, we show that SPD remains effective even without complete logit access in GPT-3.5 and GPT-4. We believe that our proposed method offers a promising approach to efficiently safeguard LLMs against adversarial attacks.

Elias Abad Rocamora, Yongtao Wu, Fanghui Liu et al.

Adversarial attacks in Natural Language Processing apply perturbations in the character or token levels. Token-level attacks, gaining prominence for their use of gradient-based methods, are susceptible to altering sentence semantics, leading to invalid adversarial examples. While character-level attacks easily maintain semantics, they have received less attention as they cannot easily adopt popular gradient-based methods, and are thought to be easy to defend. Challenging these beliefs, we introduce Charmer, an efficient query-based adversarial attack capable of achieving high attack success rate (ASR) while generating highly similar adversarial examples. Our method successfully targets both small (BERT) and large (Llama 2) models. Specifically, on BERT with SST-2, Charmer improves the ASR in 4.84% points and the USE similarity in 8% points with respect to the previous art. Our implementation is available in https://github.com/LIONS-EPFL/Charmer.

Shuibai Zhang, Fred Zhangzhi Peng, Yiheng Zhang et al.

While Diffusion Language Models (DLMs) are theoretically well-suited for iterative refinement due to their non-causal structure, they often fail to reliably revise incorrect tokens in practice. The key challenge lies in the model's inability to distinguish between correct and erroneous tokens in a visible sequence. Standard masked diffusion language model (MDLM) training is restricted to the objective of unmasking, undermining the effectiveness of refinement guided by confidence. Based on this observation, we study corrective behavior in DLMs, defined as the ability to assign lower confidence to incorrect tokens and iteratively refine them while preserving correct content. We show that this capability is not induced by conventional masked diffusion objectives and propose a post-training principle oriented by correction that explicitly supervises visible incorrect tokens, enabling discriminative confidence and targeted refinement. To evaluate corrective behavior, we introduce the Code Revision Benchmark, a controllable and executable benchmark for assessing error localization and in-place correction. Experiments on code revision tasks and parallel decoding scenarios demonstrate that models trained with our approach substantially outperform standard MDLMs, with gains that are most pronounced when parallel decoding introduces substantial uncertainty and iterative refinement becomes essential. Our code is publicly available at https://github.com/zhangshuibai/CDLM.

Elias Abad Rocamora, Fanghui Liu, Grigorios G. Chrysos et al.

Catastrophic overfitting (CO) in single-step adversarial training (AT) results in abrupt drops in the adversarial test accuracy (even down to 0%). For models trained with multi-step AT, it has been observed that the loss function behaves locally linearly with respect to the input, this is however lost in single-step AT. To address CO in single-step AT, several methods have been proposed to enforce local linearity of the loss via regularization. However, these regularization terms considerably slow down training due to Double Backpropagation. Instead, in this work, we introduce a regularization term, called ELLE, to mitigate CO effectively and efficiently in classical AT evaluations, as well as some more difficult regimes, e.g., large adversarial perturbations and long training schedules. Our regularization term can be theoretically linked to curvature of the loss function and is computationally cheaper than previous methods by avoiding Double Backpropagation. Our thorough experimental validation demonstrates that our work does not suffer from CO, even in challenging settings where previous works suffer from it. We also notice that adapting our regularization parameter during training (ELLE-A) greatly improves the performance, specially in large $ε$ setups. Our implementation is available in https://github.com/LIONS-EPFL/ELLE .

Aleksandr Timofeev, Grigorios G. Chrysos, Volkan Cevher

Neural Architecture Search (NAS) provides state-of-the-art results when trained on well-curated datasets with annotated labels. However, annotating data or even having balanced number of samples can be a luxury for practitioners from different scientific fields, e.g., in the medical domain. To that end, we propose a NAS-based framework that bears the threefold contributions: (a) we focus on the self-supervised scenario, i.e., where no labels are required to determine the architecture, and (b) we assume the datasets are imbalanced, (c) we design each component to be able to run on a resource constrained setup, i.e., on a single GPU (e.g. Google Colab). Our components build on top of recent developments in self-supervised learning~\citep{zbontar2021barlow}, self-supervised NAS~\citep{kaplan2020self} and extend them for the case of imbalanced datasets. We conduct experiments on an (artificially) imbalanced version of CIFAR-10 and we demonstrate our proposed method outperforms standard neural networks, while using $27\times$ less parameters. To validate our assumption on a naturally imbalanced dataset, we also conduct experiments on ChestMNIST and COVID-19 X-ray. The results demonstrate how the proposed method can be used in imbalanced datasets, while it can be fully run on a single GPU. Code is available \href{https://github.com/TimofeevAlex/ssnas_imbalanced}{here}.

Muhammad H. Ashiq, Samanyu Arora, Abhinav N. Harish et al.

We theoretically study the hallucination phenomena in two canonical diffusion samplers: the stochastic Denoising Diffusion Probabilistic Model (DDPM) and the deterministic Denoising Diffusion Implicit Model (DDIM). We analyze the reverse ODE (DDIM) and SDE (DDPM) for a Gaussian mixture target, proving that after a critical time $τ$, (a) DDIM can become stuck on the segment connecting the two nearest modes and (b) DDPM *stochasticity* helps it become unstuck from this region, thus avoiding hallucination. Our empirical validation verifies that DDPM has a significantly lower hallucination rate than DDIM when this region is entered. Building on our observations, we exhibit how using additional stochastic steps can help DDIM avoid hallucinations and offer new insights on how to design improved samplers.

Yixin Cheng, Markos Georgopoulos, Volkan Cevher et al.

Large Language Models (LLMs) are susceptible to Jailbreaking attacks, which aim to extract harmful information by subtly modifying the attack query. As defense mechanisms evolve, directly obtaining harmful information becomes increasingly challenging for Jailbreaking attacks. In this work, inspired from Chomsky's transformational-generative grammar theory and human practices of indirect context to elicit harmful information, we focus on a new attack form, called Contextual Interaction Attack. We contend that the prior context\u2014the information preceding the attack query\u2014plays a pivotal role in enabling strong Jailbreaking attacks. Specifically, we propose a first multi-turn approach that leverages benign preliminary questions to interact with the LLM. Due to the autoregressive nature of LLMs, which use previous conversation rounds as context during generation, we guide the model's question-response pair to construct a context that is semantically aligned with the attack query to execute the attack. We conduct experiments on seven different LLMs and demonstrate the efficacy of this attack, which is black-box and can also transfer across LLMs. We believe this can lead to further developments and understanding of security in LLMs.

Yixin Cheng, Grigorios G. Chrysos, Markos Georgopoulos et al.

Despite the remarkable capabilities of deep neural networks in image recognition, the dependence on activation functions remains a largely unexplored area and has yet to be eliminated. On the other hand, Polynomial Networks is a class of models that does not require activation functions, but have yet to perform on par with modern architectures. In this work, we aim close this gap and propose MONet, which relies solely on multilinear operators. The core layer of MONet, called Mu-Layer, captures multiplicative interactions of the elements of the input token. MONet captures high-degree interactions of the input elements and we demonstrate the efficacy of our approach on a series of image recognition and scientific computing benchmarks. The proposed model outperforms prior polynomial networks and performs on par with modern architectures. We believe that MONet can inspire further research on models that use entirely multilinear operations.

Aggelina Chatziagapi, Grigorios G. Chrysos, Dimitris Samaras

In this work, we introduce a method that learns a single dynamic neural radiance field (NeRF) from monocular talking face videos of multiple identities. NeRFs have shown remarkable results in modeling the 4D dynamics and appearance of human faces. However, they require per-identity optimization. Although recent approaches have proposed techniques to reduce the training and rendering time, increasing the number of identities can be expensive. We introduce MI-NeRF (multi-identity NeRF), a single unified network that models complex non-rigid facial motion for multiple identities, using only monocular videos of arbitrary length. The core premise in our method is to learn the non-linear interactions between identity and non-identity specific information with a multiplicative module. By training on multiple videos simultaneously, MI-NeRF not only reduces the total training time compared to standard single-identity NeRFs, but also demonstrates robustness in synthesizing novel expressions for any input identity. We present results for both facial expression transfer and talking face video synthesis. Our method can be further personalized for a target identity given only a short video.

Kostas Triaridis, Alexandros Graikos, Aggelina Chatziagapi et al.

Diffusion models, despite their impressive demos, often produce hallucinatory samples with structural inconsistencies that lie outside of the support of the true data distribution. Such hallucinations can be attributed to excessive smoothing between modes of the data distribution. However, semantic interpolations are often desirable and can lead to generation diversity, thus we believe a more nuanced solution is required. In this work, we introduce Dynamic Guidance, which tackles this issue. Dynamic Guidance mitigates hallucinations by selectively sharpening the score function only along the pre-determined directions known to cause artifacts, while preserving valid semantic variations. To our knowledge, this is the first approach that addresses hallucinations at generation time rather than through post-hoc filtering. Dynamic Guidance substantially reduces hallucinations on both controlled and natural image datasets, significantly outperforming baselines.

Elias Abad Rocamora, Grigorios G. Chrysos, Volkan Cevher

Text classifiers suffer from small perturbations, that if chosen adversarially, can dramatically change the output of the model. Verification methods can provide robustness certificates against such adversarial perturbations, by computing a sound lower bound on the robust accuracy. Nevertheless, existing verification methods incur in prohibitive costs and cannot practically handle Levenshtein distance constraints. We propose the first method for computing the Lipschitz constant of convolutional classifiers with respect to the Levenshtein distance. We use these Lipschitz constant estimates for training 1-Lipschitz classifiers. This enables computing the certified radius of a classifier in a single forward pass. Our method, LipsLev, is able to obtain $38.80$% and $13.93$% verified accuracy at distance $1$ and $2$ respectively in the AG-News dataset, while being $4$ orders of magnitude faster than existing approaches. We believe our work can open the door to more efficient verification in the text domain.

Yihang Chen, Fanghui Liu, Yiping Lu et al.

Despite the widespread empirical success of ResNet, the generalization properties of deep ResNet are rarely explored beyond the lazy training regime. In this work, we investigate \emph{scaled} ResNet in the limit of infinitely deep and wide neural networks, of which the gradient flow is described by a partial differential equation in the large-neural network limit, i.e., the \emph{mean-field} regime. To derive the generalization bounds under this setting, our analysis necessitates a shift from the conventional time-invariant Gram matrix employed in the lazy training regime to a time-variant, distribution-dependent version. To this end, we provide a global lower bound on the minimum eigenvalue of the Gram matrix under the mean-field regime. Besides, for the traceability of the dynamic of Kullback-Leibler (KL) divergence, we establish the linear convergence of the empirical error and estimate the upper bound of the KL divergence over parameters distribution. Finally, we build the uniform convergence for generalization bound via Rademacher complexity. Our results offer new insights into the generalization ability of deep ResNet beyond the lazy training regime and contribute to advancing the understanding of the fundamental properties of deep neural networks.

Yannis Panagakis, Jean Kossaifi, Grigorios G. Chrysos et al.

Tensors, or multidimensional arrays, are data structures that can naturally represent visual data of multiple dimensions. Inherently able to efficiently capture structured, latent semantic spaces and high-order interactions, tensors have a long history of applications in a wide span of computer vision problems. With the advent of the deep learning paradigm shift in computer vision, tensors have become even more fundamental. Indeed, essential ingredients in modern deep learning architectures, such as convolutions and attention mechanisms, can readily be considered as tensor mappings. In effect, tensor methods are increasingly finding significant applications in deep learning, including the design of memory and compute efficient network architectures, improving robustness to random noise and adversarial attacks, and aiding the theoretical understanding of deep networks. This article provides an in-depth and practical review of tensors and tensor methods in the context of representation learning and deep learning, with a particular focus on visual data analysis and computer vision applications. Concretely, besides fundamental work in tensor-based visual data analysis methods, we focus on recent developments that have brought on a gradual increase of tensor methods, especially in deep learning architectures, and their implications in computer vision applications. To further enable the newcomer to grasp such concepts quickly, we provide companion Python notebooks, covering key aspects of the paper and implementing them, step-by-step with TensorLy.

Grigorios G. Chrysos, Stylianos Moschoglou, Giorgos Bouritsas et al.

Deep Convolutional Neural Networks (DCNNs) is currently the method of choice both for generative, as well as for discriminative learning in computer vision and machine learning. The success of DCNNs can be attributed to the careful selection of their building blocks (e.g., residual blocks, rectifiers, sophisticated normalization schemes, to mention but a few). In this paper, we propose $Π$-Nets, a new class of DCNNs. $Π$-Nets are polynomial neural networks, i.e., the output is a high-order polynomial of the input. $Π$-Nets can be implemented using special kind of skip connections and their parameters can be represented via high-order tensors. We empirically demonstrate that $Π$-Nets have better representation power than standard DCNNs and they even produce good results without the use of non-linear activation functions in a large battery of tasks and signals, i.e., images, graphs, and audio. When used in conjunction with activation functions, $Π$-Nets produce state-of-the-art results in challenging tasks, such as image generation. Lastly, our framework elucidates why recent generative models, such as StyleGAN, improve upon their predecessors, e.g., ProGAN.

Ioannis Marras, Grigorios G. Chrysos, Ioannis Alexiou et al.

Deep Convolutional Neural Networks (CNNs) have been successfully used in many low-level vision problems like image denoising. Although the conditional image generation techniques have led to large improvements in this task, there has been little effort in providing conditional generative adversarial networks (cGAN)[42] with an explicit way of understanding the image noise for object-independent denoising reliable for real-world applications. The task of leveraging structures in the target space is unstable due to the complexity of patterns in natural scenes, so the presence of unnatural artifacts or over-smoothed image areas cannot be avoided. To fill the gap, in this work we introduce the idea of a cGAN which explicitly leverages structure in the image noise space. By learning directly a low dimensional manifold of the image noise, the generator promotes the removal from the noisy image only that information which spans this manifold. This idea brings many advantages while it can be appended at the end of any denoiser to significantly improve its performance. Based on our experiments, our model substantially outperforms existing state-of-the-art architectures, resulting in denoised images with less oversmoothing and better detail.

Grigorios G. Chrysos, Jean Kossaifi, Stefanos Zafeiriou

Conditional generative adversarial networks (cGAN) have led to large improvements in the task of conditional image generation, which lies at the heart of computer vision. The major focus so far has been on performance improvement, while there has been little effort in making cGAN more robust to noise. The regression (of the generator) might lead to arbitrarily large errors in the output, which makes cGAN unreliable for real-world applications. In this work, we introduce a novel conditional GAN model, called RoCGAN, which leverages structure in the target space of the model to address the issue. Our model augments the generator with an unsupervised pathway, which promotes the outputs of the generator to span the target manifold even in the presence of intense noise. We prove that RoCGAN share similar theoretical properties as GAN and experimentally verify that our model outperforms existing state-of-the-art cGAN architectures by a large margin in a variety of domains including images from natural scenes and faces.

Grigorios G. Chrysos, Paolo Favaro, Stefanos Zafeiriou

Face analysis is a core part of computer vision, in which remarkable progress has been observed in the past decades. Current methods achieve recognition and tracking with invariance to fundamental modes of variation such as illumination, 3D pose, expressions. Notwithstanding, a much less standing mode of variation is motion deblurring, which however presents substantial challenges in face analysis. Recent approaches either make oversimplifying assumptions, e.g. in cases of joint optimization with other tasks, or fail to preserve the highly structured shape/identity information. Therefore, we propose a data-driven method that encourages identity preservation. The proposed model includes two parallel streams (sub-networks): the first deblurs the image, the second implicitly extracts and projects the identity of both the sharp and the blurred image in similar subspaces. We devise a method for creating realistic motion blur by averaging a variable number of frames to train our model. The averaged images originate from a 2MF2 dataset with 10 million facial frames, which we introduce for the task. Considering deblurring as an intermediate step, we utilize the deblurred outputs to conduct a thorough experimentation on high-level face analysis tasks, i.e. landmark localization and face verification. The experimental evaluation demonstrates the superiority of our method.

Grigorios G. Chrysos, Yannis Panagakis, Stefanos Zafeiriou

The rapid progress in machine learning methods has been empowered by i) huge datasets that have been collected and annotated, ii) improved engineering (e.g. data pre-processing/normalization). The existing datasets typically include several million samples, which constitutes their extension a colossal task. In addition, the state-of-the-art data-driven methods demand a vast amount of data, hence a standard engineering trick employed is artificial data augmentation for instance by adding into the data cropped and (affinely) transformed images. However, this approach does not correspond to any change in the natural 3D scene. We propose instead to perform data augmentation through learning realistic local transformations. We learn a forward and an inverse transformation that maps an image from the high-dimensional space of pixel intensities to a latent space which varies (approximately) linearly with the latent space of a realistically transformed version of the image. Such transformed images can be considered two successive frames in a video. Next, we utilize these transformations to learn a linear model that modifies the latent spaces and then use the inverse transformation to synthesize a new image. We argue that the this procedure produces powerful invariant representations. We perform both qualitative and quantitative experiments that demonstrate our proposed method creates new realistic images.

Grigorios G. Chrysos, Stefanos Zafeiriou

Blind deblurring consists a long studied task, however the outcomes of generic methods are not effective in real world blurred images. Domain-specific methods for deblurring targeted object categories, e.g. text or faces, frequently outperform their generic counterparts, hence they are attracting an increasing amount of attention. In this work, we develop such a domain-specific method to tackle deblurring of human faces, henceforth referred to as face deblurring. Studying faces is of tremendous significance in computer vision, however face deblurring has yet to demonstrate some convincing results. This can be partly attributed to the combination of i) poor texture and ii) highly structure shape that yield the contour/gradient priors (that are typically used) sub-optimal. In our work instead of making assumptions over the prior, we adopt a learning approach by inserting weak supervision that exploits the well-documented structure of the face. Namely, we utilise a deep network to perform the deblurring and employ a face alignment technique to pre-process each face. We additionally surpass the requirement of the deep network for thousands training samples, by introducing an efficient framework that allows the generation of a large dataset. We utilised this framework to create 2MF2, a dataset of over two million frames. We conducted experiments with real world blurred facial images and report that our method returns a result close to the sharp natural latent image.

Grigorios G. Chrysos, Epameinondas Antonakos, Patrick Snape et al.

Recently, technologies such as face detection, facial landmark localisation and face recognition and verification have matured enough to provide effective and efficient solutions for imagery captured under arbitrary conditions (referred to as "in-the-wild"). This is partially attributed to the fact that comprehensive "in-the-wild" benchmarks have been developed for face detection, landmark localisation and recognition/verification. A very important technology that has not been thoroughly evaluated yet is deformable face tracking "in-the-wild". Until now, the performance has mainly been assessed qualitatively by visually assessing the result of a deformable face tracking technology on short videos. In this paper, we perform the first, to the best of our knowledge, thorough evaluation of state-of-the-art deformable face tracking pipelines using the recently introduced 300VW benchmark. We evaluate many different architectures focusing mainly on the task of on-line deformable face tracking. In particular, we compare the following general strategies: (a) generic face detection plus generic facial landmark localisation, (b) generic model free tracking plus generic facial landmark localisation, as well as (c) hybrid approaches using state-of-the-art face detection, model free tracking and facial landmark localisation technologies. Our evaluation reveals future avenues for further research on the topic.