Cristina Nader Vasconcelos

Mostafa Dehghani, Josip Djolonga, Basil Mustafa et al. · deepmind

The scaling of Transformers has driven breakthrough capabilities for language models. At present, the largest large language models (LLMs) contain upwards of 100B parameters. Vision Transformers (ViT) have introduced the same architecture to image and video modelling, but these have not yet been successfully scaled to nearly the same degree; the largest dense ViT contains 4B parameters (Chen et al., 2022). We present a recipe for highly efficient and stable training of a 22B-parameter ViT (ViT-22B) and perform a wide variety of experiments on the resulting model. When evaluated on downstream tasks (often with a lightweight linear model on frozen features), ViT-22B demonstrates increasing performance with scale. We further observe other interesting benefits of scale, including an improved tradeoff between fairness and performance, state-of-the-art alignment to human visual perception in terms of shape/texture bias, and improved robustness. ViT-22B demonstrates the potential for "LLM-like" scaling in vision, and provides key steps towards getting there.

Cristina Vasconcelos, Cengiz Oztireli, Mark Matthews et al. · deepmind

Neural fields have rapidly been adopted for representing 3D signals, but their application to more classical 2D image-processing has been relatively limited. In this paper, we consider one of the most important operations in image processing: upsampling. In deep learning, learnable upsampling layers have extensively been used for single image super-resolution. We propose to parameterize upsampling kernels as neural fields. This parameterization leads to a compact architecture that obtains a 40-fold reduction in the number of parameters when compared with competing arbitrary-scale super-resolution architectures. When upsampling images of size 256x256 we show that our architecture is 2x-10x more efficient than competing arbitrary-scale super-resolution architectures, and more efficient than sub-pixel convolutions when instantiated to a single-scale model. In the general setting, these gains grow polynomially with the square of the target scale. We validate our method on standard benchmarks showing such efficiency gains can be achieved without sacrifices in super-resolution performance.

Cristina Vasconcelos, Vighnesh Birodkar, Vincent Dumoulin

A common practice in transfer learning is to initialize the downstream model weights by pre-training on a data-abundant upstream task. In object detection specifically, the feature backbone is typically initialized with Imagenet classifier weights and fine-tuned on the object detection task. Recent works show this is not strictly necessary under longer training regimes and provide recipes for training the backbone from scratch. We investigate the opposite direction of this end-to-end training trend: we show that an extreme form of knowledge preservation -- freezing the classifier-initialized backbone -- consistently improves many different detection models, and leads to considerable resource savings. We hypothesize and corroborate experimentally that the remaining detector components capacity and structure is a crucial factor in leveraging the frozen backbone. Immediate applications of our findings include performance improvements on hard cases like detection of long-tail object classes and computational and memory resource savings that contribute to making the field more accessible to researchers with access to fewer computational resources.

Zhilin Guo, Jing Yang, Kyle Fogarty et al.

Digitizing humans and synthesizing photorealistic avatars with explicit 3D pose and camera controls are central to VR, telepresence, and entertainment. Existing skinning-based workflows require laborious manual rigging or template-based fittings, while neural volumetric methods rely on canonical templates and re-optimization for each unseen pose. We present PoseCraft, a diffusion framework built around tokenized 3D interface: instead of relying only on rasterized geometry as 2D control images, we encode sparse 3D landmarks and camera extrinsics as discrete conditioning tokens and inject them into diffusion via cross-attention. Our approach preserves 3D semantics by avoiding 2D re-projection ambiguity under large pose and viewpoint changes, and produces photorealistic imagery that faithfully captures identity and appearance. To train and evaluate at scale, we also implement GenHumanRF, a data generation workflow that renders diverse supervision from volumetric reconstructions. Our experiments show that PoseCraft achieves significant perceptual quality improvement over diffusion-centric methods, and attains better or comparable metrics to latest volumetric rendering SOTA while better preserving fabric and hair details.

Cristina Vasconcelos, Hugo Larochelle, Vincent Dumoulin et al.

We investigate the impact of aliasing on generalization in Deep Convolutional Networks and show that data augmentation schemes alone are unable to prevent it due to structural limitations in widely used architectures. Drawing insights from frequency analysis theory, we take a closer look at ResNet and EfficientNet architectures and review the trade-off between aliasing and information loss in each of their major components. We show how to mitigate aliasing by inserting non-trainable low-pass filters at key locations, particularly where networks lack the capacity to learn them. These simple architectural changes lead to substantial improvements in generalization on i.i.d. and even more on out-of-distribution conditions, such as image classification under natural corruptions on ImageNet-C [11] and few-shot learning on Meta-Dataset [26]. State-of-the art results are achieved on both datasets without introducing additional trainable parameters and using the default hyper-parameters of open source codebases.

Cristina Vasconcelos, Hugo Larochelle, Vincent Dumoulin et al.

Image pre-processing in the frequency domain has traditionally played a vital role in computer vision and was even part of the standard pipeline in the early days of deep learning. However, with the advent of large datasets, many practitioners concluded that this was unnecessary due to the belief that these priors can be learned from the data itself. Frequency aliasing is a phenomenon that may occur when sub-sampling any signal, such as an image or feature map, causing distortion in the sub-sampled output. We show that we can mitigate this effect by placing non-trainable blur filters and using smooth activation functions at key locations, particularly where networks lack the capacity to learn them. These simple architectural changes lead to substantial improvements in out-of-distribution generalization on both image classification under natural corruptions on ImageNet-C [10] and few-shot learning on Meta-Dataset [17], without introducing additional trainable parameters and using the default hyper-parameters of open source codebases.

Xingchang Huang, Corentin Salaün, Cristina Vasconcelos et al.

Most of the existing diffusion models use Gaussian noise for training and sampling across all time steps, which may not optimally account for the frequency contents reconstructed by the denoising network. Despite the diverse applications of correlated noise in computer graphics, its potential for improving the training process has been underexplored. In this paper, we introduce a novel and general class of diffusion models taking correlated noise within and across images into account. More specifically, we propose a time-varying noise model to incorporate correlated noise into the training process, as well as a method for fast generation of correlated noise mask. Our model is built upon deterministic diffusion models and utilizes blue noise to help improve the generation quality compared to using Gaussian white (random) noise only. Further, our framework allows introducing correlation across images within a single mini-batch to improve gradient flow. We perform both qualitative and quantitative evaluations on a variety of datasets using our method, achieving improvements on different tasks over existing deterministic diffusion models in terms of FID metric.

Luis Miguel Pazos-Outón, Cristina Nader Vasconcelos, Anton Raichuk et al.

Protecting and restoring forest ecosystems is critical for biodiversity conservation and carbon sequestration. Forest monitoring on a global scale is essential for prioritizing and assessing conservation efforts. Satellite-based remote sensing is the only viable solution for providing global coverage, but to date, large-scale forest monitoring is limited to single modalities and single time points. In this paper, we present a dataset consisting of data from five public satellites for recognizing forest plantations and planted tree species across the globe. Each satellite modality consists of a multi-year time series. The dataset, named \PlantD, includes over 2M examples of 64 tree label classes (46 genera and 40 species), distributed among 41 countries. This dataset is released to foster research in forest monitoring using multimodal, multi-scale, multi-temporal data sources. Additionally, we present initial baseline results and evaluate modality fusion and data augmentation approaches for this dataset.

Ibtihel Amara, Ahmed Imtiaz Humayun, Ivana Kajic et al.

Concept erasure techniques have recently gained significant attention for their potential to remove unwanted concepts from text-to-image models. While these methods often demonstrate promising results in controlled settings, their robustness in real-world applications and suitability for deployment remain uncertain. In this work, we (1) identify a critical gap in evaluating sanitized models, particularly in assessing their performance across diverse concept dimensions, and (2) systematically analyze the failure modes of text-to-image models post-erasure. We focus on the unintended consequences of concept removal on non-target concepts across different levels of interconnected relationships including visually similar, binomial, and semantically related concepts. To address this, we introduce EraseBench, a comprehensive benchmark for evaluating post-erasure performance. EraseBench includes over 100 curated concepts, targeted evaluation prompts, and a robust set of metrics to assess both effectiveness and side effects of erasure. Our findings reveal a phenomenon of concept entanglement, where erasure leads to unintended suppression of non-target concepts, causing spillover degradation that manifests as distortions and a decline in generation quality.

Eduardo Vera Sousa, Leandro A. F. Fernandes, Cristina Nader Vasconcelos

Convolutional Neural Networks (CNNs) have been widely applied. But as the CNNs grow, the number of arithmetic operations and memory footprint also increase. Furthermore, typical non-linear activation functions do not allow associativity of the operations encoded by consecutive layers, preventing the simplification of intermediate steps by combining them. We present a new activation function that allows associativity between sequential layers of CNNs. Even though our activation function is non-linear, it can be represented by a sequence of linear operations in the conformal model for Euclidean geometry. In this domain, operations like, but not limited to, convolution, average pooling, and dropout remain linear. We take advantage of associativity to combine all the "conformal layers" and make the cost of inference constant regardless of the depth of the network.

Bruno Augusto Dorta Marques, Esteban Walter Gonzalez Clua, Anselmo Antunes Montenegro et al.

The representation of consistent mixed reality (XR) environments requires adequate real and virtual illumination composition in real-time. Estimating the lighting of a real scenario is still a challenge. Due to the ill-posed nature of the problem, classical inverse-rendering techniques tackle the problem for simple lighting setups. However, those assumptions do not satisfy the current state-of-art in computer graphics and XR applications. While many recent works solve the problem using machine learning techniques to estimate the environment light and scene's materials, most of them are limited to geometry or previous knowledge. This paper presents a CNN-based model to estimate complex lighting for mixed reality environments with no previous information about the scene. We model the environment illumination using a set of spherical harmonics (SH) environment lighting, capable of efficiently represent area lighting. We propose a new CNN architecture that inputs an RGB image and recognizes, in real-time, the environment lighting. Unlike previous CNN-based lighting estimation methods, we propose using a highly optimized deep neural network architecture, with a reduced number of parameters, that can learn high complex lighting scenarios from real-world high-dynamic-range (HDR) environment images. We show in the experiments that the CNN architecture can predict the environment lighting with an average mean squared error (MSE) of \num{7.85e-04} when comparing SH lighting coefficients. We validate our model in a variety of mixed reality scenarios. Furthermore, we present qualitative results comparing relights of real-world scenes.

Fartash Faghri, Sven Gowal, Cristina Vasconcelos et al.

We demonstrate that the choice of optimizer, neural network architecture, and regularizer significantly affect the adversarial robustness of linear neural networks, providing guarantees without the need for adversarial training. To this end, we revisit a known result linking maximally robust classifiers and minimum norm solutions, and combine it with recent results on the implicit bias of optimizers. First, we show that, under certain conditions, it is possible to achieve both perfect standard accuracy and a certain degree of robustness, simply by training an overparametrized model using the implicit bias of the optimization. In that regime, there is a direct relationship between the type of the optimizer and the attack to which the model is robust. To the best of our knowledge, this work is the first to study the impact of optimization methods such as sign gradient descent and proximal methods on adversarial robustness. Second, we characterize the robustness of linear convolutional models, showing that they resist attacks subject to a constraint on the Fourier-$\ell_\infty$ norm. To illustrate these findings we design a novel Fourier-$\ell_\infty$ attack that finds adversarial examples with controllable frequencies. We evaluate Fourier-$\ell_\infty$ robustness of adversarially-trained deep CIFAR-10 models from the standard RobustBench benchmark and visualize adversarial perturbations.

Fábio Perez, Cristina Vasconcelos, Sandra Avila et al.

Deep learning models show remarkable results in automated skin lesion analysis. However, these models demand considerable amounts of data, while the availability of annotated skin lesion images is often limited. Data augmentation can expand the training dataset by transforming input images. In this work, we investigate the impact of 13 data augmentation scenarios for melanoma classification trained on three CNNs (Inception-v4, ResNet, and DenseNet). Scenarios include traditional color and geometric transforms, and more unusual augmentations such as elastic transforms, random erasing and a novel augmentation that mixes different lesions. We also explore the use of data augmentation at test-time and the impact of data augmentation on various dataset sizes. Our results confirm the importance of data augmentation in both training and testing and show that it can lead to more performance gains than obtaining new images. The best scenario results in an AUC of 0.882 for melanoma classification without using external data, outperforming the top-ranked submission (0.874) for the ISIC Challenge 2017, which was trained with additional data.

Cristina Nader Vasconcelos, Bárbara Nader Vasconcelos

Skin cancer is a major public health problem, as is the most common type of cancer and represents more than half of cancer diagnoses worldwide. Early detection influences the outcome of the disease and motivates our work. We investigate the composition of CNN committees and data augmentation for the the ISBI 2017 Melanoma Classification Challenge (named Skin Lesion Analysis towards Melanoma Detection) facing the peculiarities of dealing with such a small, unbalanced, biological database. For that, we explore committees of Convolutional Neural Networks trained over the ISBI challenge training dataset artificially augmented by both classical image processing transforms and image warping guided by specialist knowledge about the lesion axis and improve the final classifier invariance to common melanoma variations.

Thiago M. Porcino, Esteban W. Clua, Cristina N. Vasconcelos et al.

We are experiencing an upcoming trend of using head mounted display systems in games and serious games, which is likely to become an established practice in the near future. While these systems provide highly immersive experiences, many users have been reporting discomfort symptoms, such as nausea, sickness, and headaches, among others. When using VR for health applications, this is more critical, since the discomfort may interfere a lot in treatments. In this work we discuss possible causes of these issues, and present possible solutions as design guidelines that may mitigate them. In this context, we go deeper within a dynamic focus solution to reduce discomfort in immersive virtual environments, when using first-person navigation. This solution applies an heuristic model of visual attention that works in real time. This work also discusses a case study (as a first-person spatial shooter demo) that applies this solution and the proposed design guidelines.