NeurIPS 2024 Papers

In recent years, the integration of vision and language understanding has led to significant advancements in artificial intelligence, particularly through Vision-Language Models (VLMs). However, existing VLMs face challenges in handling real-world applications with complex scenes and multiple objects, as well as aligning their focus with the diverse attention patterns of human users. In this paper, we introduce gaze information, feasibly collected by ubiquitous wearable devices such as MR glasses, as a proxy for human attention to guide VLMs. We propose a novel approach, Voila-A, for gaze alignment to enhance the effectiveness of these models in real-world applications. First, we collect hundreds of minutes of gaze data to demonstrate that we can mimic human gaze modalities using localized narratives. We then design an automatic data annotation pipeline utilizing GPT-4 to generate the VOILA-COCO dataset. Additionally, we introduce a new model VOILA-A that integrate gaze information into VLMs while maintain pretrained knowledge from webscale dataset. We evaluate Voila-A using a hold-out validation set and a newly collected VOILA-GAZE testset, which features real-life scenarios captured with a gaze-tracking device. Our experimental results demonstrate that Voila-A significantly outperforms several baseline models. By aligning model attention with human gaze patterns, Voila-A paves the way for more intuitive, user-centric VLMs and fosters engaging human-AI interaction across a wide range of applications.

Attribution maps are one of the most established tools to explain the functioning of computer vision models. They assign importance scores to input features, indicating how relevant each feature is for the prediction of a deep neural network. While much research has gone into proposing new attribution methods, their proper evaluation remains a difficult challenge. In this work, we propose a novel evaluation protocol that overcomes two fundamental limitations of the widely used incremental-deletion protocol, i.e., the out-of-domain issue and lacking inter-model comparisons. This allows us to evaluate 23 attribution methods and how different design choices of popular vision backbones affect their attribution quality. We find that intrinsically explainable models outperform standard models and that raw attribution values exhibit a higher attribution quality than what is known from previous work. Further, we show consistent changes in the attribution quality when varying the network design, indicating that some standard design choices promote attribution quality.

High-throughput image analysis in the biomedical domain has gained significant attention in recent years, driving advancements in drug discovery, disease prediction, and personalized medicine. Organoids, specifically, are an active area of research, providing excellent models for human organs and their functions. Automating the quantification of organoids in microscopy images would provide an effective solution to overcome substantial manual quantification bottlenecks, particularly in high-throughput image analysis. However, there is a notable lack of open biomedical datasets, in contrast to other domains, such as autonomous driving, and, notably, only few of them have attempted to quantify annotation uncertainty. In this work, we present MultiOrg a comprehensive organoid dataset tailored for object detection tasks with uncertainty quantification. This dataset comprises over 400 high-resolution 2d microscopy images and curated annotations of more than 60,000 organoids. Most importantly, it includes three label sets for the test data, independently annotated by two experts at distinct time points. We additionally provide a benchmark for organoid detection, and make the best model available through an easily installable, interactive plugin for the popular image visualization tool Napari, to perform organoid quantification.

Despite the importance of shape perception in human vision, early neural image classifiers relied less on shape information for object recognition than other (often spurious) features. While recent research suggests that current large Vision-Language Models (VLMs) exhibit more reliance on shape, we find them to still be seriously limited in this regard. To quantify such limitations, we introduce IllusionBench, a dataset that challenges current cutting-edge VLMs to decipher shape information when the shape is represented by an arrangement of visual elements in a scene. Our extensive evaluations reveal that, while these shapes are easily detectable by human annotators, current VLMs struggle to recognize them, indicating important avenues for future work in developing more robust visual perception systems. The full dataset and codebase are available at: https://arshiahemmat.github.io/illusionbench/

Model editing aims to data-efficiently correct predictive errors of large pre-trained models while ensuring generalization to neighboring failures and locality to minimize unintended effects on unrelated examples. While significant progress has been made in editing Transformer-based large language models, effective strategies for editing vision Transformers (ViTs) in computer vision remain largely untapped. In this paper, we take initial steps towards correcting predictive errors of ViTs, particularly those arising from subpopulation shifts. Taking a locate-then-edit approach, we first address the ``where-to-edit`` challenge by meta-learning a hypernetwork on CutMix-augmented data generated for editing reliability. This trained hypernetwork produces generalizable binary masks that identify a sparse subset of structured model parameters, responsive to real-world failure samples. Afterward, we solve the ``how-to-edit`` problem by simply fine-tuning the identified parameters using a variant of gradient descent to achieve successful edits. To validate our method, we construct an editing benchmark that introduces subpopulation shifts towards natural underrepresented images and AI-generated images, thereby revealing the limitations of pre-trained ViTs for object recognition. Our approach not only achieves superior performance on the proposed benchmark but also allows for adjustable trade-offs between generalization and locality. Our code is available at https://github.com/hustyyq/Where-to-Edit.

Vision transformer model (ViT) is widely used and performs well in vision tasks due to its ability to capture long-range dependencies. However, the time complexity and memory consumption increase quadratically with the number of input patches which limits the usage of ViT in real-world applications. Previous methods have employed linear attention to mitigate the complexity of the original self-attention mechanism at the expense of effectiveness. In this paper, we propose QT-ViT models that improve the previous linear self-attention using quadratic Taylor expansion. Specifically, we substitute the softmax-based attention with second-order Taylor expansion, and then accelerate the quadratic expansion by reducing the time complexity with a fast approximation algorithm. The proposed method capitalizes on the property of quadratic expansion to achieve superior performance while employing linear approximation for fast inference. Compared to previous studies of linear attention, our approach does not necessitate knowledge distillation or high-order attention residuals to facilitate the training process. Extensive experiments demonstrate the efficiency and effectiveness of the proposed QT-ViTs, showcasing the state-of-the-art results. Particularly, the proposed QT-ViTs consistently surpass the previous SOTA EfficientViTs under different model sizes, and achieve a new Pareto-front in terms of accuracy and speed.

When solving ill-posed inverse problems, one often desires to explore the space of potential solutions rather than be presented with a single plausible reconstruction. Valuable insights into these feasible solutions and their associated probabilities are embedded in the posterior distribution. However, when confronted with data of high dimensionality (such as images), visualizing this distribution becomes a formidable challenge, necessitating the application of effective summarization techniques before user examination. In this work, we introduce a new approach for visualizing posteriors across multiple levels of granularity using *tree*-valued predictions. Our method predicts a tree-valued hierarchical summarization of the posterior distribution for any input measurement, in a single forward pass of a neural network. We showcase the efficacy of our approach across diverse datasets and image restoration challenges, highlighting its prowess in uncertainty quantification and visualization. Our findings reveal that our method performs comparably to a baseline that hierarchically clusters samples from a diffusion-based posterior sampler, yet achieves this with orders of magnitude greater speed. Code and examples are available at our [webpage](https://eliasnehme.github.io/PosteriorTrees/).

We introduce a new benchmark, TAPVid-3D, for evaluating the task of long-range Tracking Any Point in 3D (TAP-3D). While point tracking in two dimensions (TAP-2D) has many benchmarks measuring performance on real-world videos, such as TAPVid-DAVIS, three-dimensional point tracking has none. To this end, leveraging existing footage, we build a new benchmark for 3D point tracking featuring 4,000+ real-world videos, composed of three different data sources spanning a variety of object types, motion patterns, and indoor and outdoor environments. To measure performance on the TAP-3D task, we formulate a collection of metrics that extend the Jaccard-based metric used in TAP-2D to handle the complexities of ambiguous depth scales across models, occlusions, and multi-track spatio-temporal smoothness. We manually verify a large sample of trajectories to ensure correct video annotations, and assess the current state of the TAP-3D task by constructing competitive baselines using existing tracking models. We anticipate this benchmark will serve as a guidepost to improve our ability to understand precise 3D motion and surface deformation from monocular video.

Vision-Language Models (VLMs) have emerged as the state-of-the-art representation learning solution, with myriads of downstream applications such as image classification, retrieval and generation. A natural question is whether these models memorize their training data, which also has implications for generalization. We propose a new method for measuring memorization in VLMs, which we call dèjá vu memorization. For VLMs trained on image-caption pairs, we show that the model indeed retains information about individual objects in the training images beyond what can be inferred from correlations or the image caption. We evaluate dèjá vu memorization at both sample and population level, and show that it is significant for OpenCLIP trained on as many as 50M image-caption pairs. Finally, we show that text randomization considerably mitigates memorization risk while only moderately impacting the model’s downstream task performance. The code is available here: https://github.com/facebookresearch/VLMDejaVu.

Internet of Things and Deep Learning are synergetically and exponentially growing industrial fields with a massive call for their unification into a common framework called Edge AI. While on-device inference is a well-explored topic in recent research, backpropagation remains an open challenge due to its prohibitive computational and memory costs compared to the extreme resource constraints of embedded devices. Drawing on tensor decomposition research, we tackle the main bottleneck of backpropagation, namely the memory footprint of activation map storage. We investigate and compare the effects of activation compression using Singular Value Decomposition and its tensor variant, High-Order Singular Value Decomposition. The application of low-order decomposition results in considerable memory savings while preserving the features essential for learning, and also offers theoretical guarantees to convergence. Experimental results obtained on main-stream architectures and tasks demonstrate Pareto-superiority over other state-of-the-art solutions, in terms of the trade-off between generalization and memory footprint.

Smoke segmentation is of great importance in precisely identifying the smoke location, enabling timely fire rescue and gas leak detection. However, due to the visual diversity and blurry edges of the non-grid smoke, noisy labels are almost inevitable in large-scale pixel-level smoke datasets. Noisy labels significantly impact the robustness of the model and may lead to serious accidents. Nevertheless, currently, there are no specific methods for addressing noisy labels in smoke segmentation. Smoke differs from regular objects as its transparency varies, causing inconsistent features in the noisy labels. In this paper, we propose a conditional sample weighting (CoSW). CoSW utilizes a multi-prototype framework, where prototypes serve as prior information to apply different weighting criteria to the different feature clusters. A novel regularized within-prototype entropy (RWE) is introduced to achieve CoSW and stable prototype update. The experiments show that our approach achieves SOTA performance on both real-world and synthetic noisy smoke segmentation datasets.

The sparsely activated mixture of experts (MoE) model presents an effective alternative to densely activated (dense) models, combining improved accuracy with computational efficiency. However, training MoE models from scratch requires extensive data and computational resources, a challenge that limits their widespread adoption. To address this, we introduce MoE Jetpack, a framework designed to fine-tune the abundant and easily accessible dense checkpoints into MoE models. MoE Jetpack incorporates two key techniques: (1) **checkpoint recycling**, which initializes MoE models with dense checkpoints to accelerate convergence and enhance accuracy, minimizing the need for extensive pre-training; (2) the **hyperspherical adaptive MoE (SpheroMoE) layer**, which optimizes the MoE architecture to enhance fine-tuning performance and efficiency. Experimental results indicate that MoE Jetpack doubles the convergence speed and enhances accuracy by 2.8% on ImageNet-1K. On smaller datasets, it achieves up to 8-fold faster convergence and over 30% accuracy gains, highlighting its efficiency. The code is available at https://github.com/Adlith/MoE-Jetpack.

We propose a new method named LoD-Loc for visual localization in the air. Unlike existing localization algorithms, LoD-Loc does not rely on complex 3D representations and can estimate the pose of an Unmanned Aerial Vehicle (UAV) using a Level-of-Detail (LoD) 3D map. LoD-Loc mainly achieves this goal by aligning the wireframe derived from the LoD projected model with that predicted by the neural network. Specifically, given a coarse pose provided by the UAV sensor, LoD-Loc hierarchically builds a cost volume for uniformly sampled pose hypotheses to describe pose probability distribution and select a pose with maximum probability. Each cost within this volume measures the degree of line alignment between projected and predicted wireframes. LoD-Loc also devises a 6-DoF pose optimization algorithm to refine the previous result with a differentiable Gaussian-Newton method. As no public dataset exists for the studied problem, we collect two datasets with map levels of LoD3.0 and LoD2.0, along with real RGB queries and ground-truth pose annotations. We benchmark our method and demonstrate that LoD-Loc achieves excellent performance, even surpassing current state-of-the-art methods that use textured 3D models for localization. The code and dataset will be made available upon publication.

This paper introduces FUNGI, **F**eatures from **UN**supervised **G**rad**I**ents, a method to enhance the features of transformer encoders by leveraging self-supervised gradients. Our method is simple: given any pretrained model, we first compute gradients from various self-supervised objectives for each input. These gradients are projected to a lower dimension and then concatenated with the model's output embedding. The resulting features are evaluated on k-nearest neighbor classification over 11 datasets from vision, 5 from natural language processing, and 2 from audio. Across backbones spanning various sizes and pretraining strategies, FUNGI features provide consistent performance improvements over the embeddings. We also show that using FUNGI features can benefit linear classification, clustering and image retrieval, and that they significantly improve the retrieval-based in-context scene understanding abilities of pretrained models, for example improving upon DINO by +17% for semantic segmentation - without any training. Code is available at https://github.com/WalterSimoncini/fungivision.

Increasing the throughput of the Transformer architecture, a foundational component used in numerous state-of-the-art models for vision and language tasks (e.g., GPT, LLaVa), is an important problem in machine learning. One recent and effective strategy is to merge token representations within Transformer models, aiming to reduce computational and memory requirements while maintaining accuracy. Prior work has proposed algorithms based on Bipartite Soft Matching (BSM), which divides tokens into distinct sets and merges the top $k$ similar tokens. However, these methods have significant drawbacks, such as sensitivity to token-splitting strategies and damage to informative tokens in later layers. This paper presents a novel paradigm called PiToMe, which prioritizes the preservation of informative tokens using an additional metric termed the \textit{energy score}. This score identifies large clusters of similar tokens as high-energy, indicating potential candidates for merging, while smaller (unique and isolated) clusters are considered as low-energy and preserved. Experimental findings demonstrate that PiToMe saved from 40-60\% FLOPs of the base models while exhibiting superior off-the-shelf performance on image classification (0.5\% average performance drop of ViT-MAEH compared to 2.6\% as baselines), image-text retrieval (0.3\% average performance drop of Clip on Flick30k compared to 4.5\% as others), and analogously in visual questions answering with LLaVa-7B. Furthermore, PiToMe is theoretically shown to preserve intrinsic spectral properties to the original token space under mild conditions.

High-level visual brain regions contain subareas in which neurons appear to respond more strongly to examples of a particular semantic category, like faces or bodies, rather than objects. However, recent work has shown that while this finding holds on average, some out-of-category stimuli also activate neurons in these regions. This may be due to visual features common among the preferred class also being present in other images. Here, we propose a deep-learning-based approach for visualizing these features. For each neuron, we identify relevant visual features driving its selectivity by modelling responses to images based on latent activations of a deep neural network. Given an out-of-category image which strongly activates the neuron, our method first identifies a reference image from the preferred category yielding a similar feature activation pattern. We then backpropagate latent activations of both images to the pixel level, while enhancing the identified shared dimensions and attenuating non-shared features. The procedure highlights image regions containing shared features driving responses of the model neuron. We apply the algorithm to novel recordings from body-selective regions in macaque IT cortex in order to understand why some images of objects excite these neurons. Visualizations reveal object parts which resemble parts of a macaque body, shedding light on neural preference of these objects.

Image pyramids are commonly used in modern computer vision tasks to obtain multi-scale features for precise understanding of images. However, image pyramids process multiple resolutions of images using the same large-scale model, which requires significant computational cost. To overcome this issue, we propose a novel network architecture known as the Parameter-Inverted Image Pyramid Networks (PIIP). Our core idea is to use models with different parameter sizes to process different resolution levels of the image pyramid, thereby balancing computational efficiency and performance. Specifically, the input to PIIP is a set of multi-scale images, where higher resolution images are processed by smaller networks. We further propose a feature interaction mechanism to allow features of different resolutions to complement each other and effectively integrate information from different spatial scales. Extensive experiments demonstrate that the PIIP achieves superior performance in tasks such as object detection, segmentation, and image classification, compared to traditional image pyramid methods and single-branch networks, while reducing computational cost. Notably, when applying our method on a large-scale vision foundation model InternViT-6B, we improve its performance by 1\%-2\% on detection and segmentation with only 40\%-60\% of the original computation. These results validate the effectiveness of the PIIP approach and provide a new technical direction for future vision computing tasks.

Concept activation vector (CAV) has attracted broad research interest in explainable AI, by elegantly attributing model predictions to specific concepts. However, the training of CAV often necessitates a large number of high-quality images, which are expensive to curate and thus limited to a predefined set of concepts. To address this issue, we propose Language-Guided CAV (LG-CAV) to harness the abundant concept knowledge within the certain pre-trained vision-language models (e.g., CLIP). This method allows training any CAV without labeled data, by utilizing the corresponding concept descriptions as guidance. To bridge the gap between vision-language model and the target model, we calculate the activation values of concept descriptions on a common pool of images (probe images) with vision-language model and utilize them as language guidance to train the LG-CAV. Furthermore, after training high-quality LG-CAVs related to all the predicted classes in the target model, we propose the activation sample reweighting (ASR), serving as a model correction technique, to improve the performance of the target model in return. Experiments on four datasets across nine architectures demonstrate that LG-CAV achieves significantly superior quality to previous CAV methods given any concept, and our model correction method achieves state-of-the-art performance compared to existing concept-based methods. Our code is available at https://github.com/hqhQAQ/LG-CAV.

Fairness in image restoration tasks is the desire to treat different sub-groups of images equally well. Existing definitions of fairness in image restoration are highly restrictive. They consider a reconstruction to be a correct outcome for a group (e.g., women) *only* if it falls within the group's set of ground truth images (e.g., natural images of women); otherwise, it is considered *entirely* incorrect. Consequently, such definitions are prone to controversy, as errors in image restoration can manifest in various ways. In this work we offer an alternative approach towards fairness in image restoration, by considering the *Group Perceptual Index* (GPI), which we define as the statistical distance between the distribution of the group's ground truth images and the distribution of their reconstructions. We assess the fairness of an algorithm by comparing the GPI of different groups, and say that it achieves perfect *Perceptual Fairness* (PF) if the GPIs of all groups are identical. We motivate and theoretically study our new notion of fairness, draw its connection to previous ones, and demonstrate its utility on state-of-the-art face image restoration algorithms.