Are you SURE? Enhancing Multimodal Pretraining with Missing Modalities through Uncertainty Estimation

Multimodal learning has demonstrated incredible successes by integrating diverse data sources, yet it often relies on the availability of all modalities - an assumption that rarely holds in real-world applications. Pretrained multimodal models, while effective, struggle when confronted with small-scale and incomplete datasets (i.e., missing modalities), limiting their practical applicability. Previous studies on reconstructing missing modalities have overlooked the reconstruction's potential unreliability, which could compromise the quality of the final outputs. We present SURE (Scalable Uncertainty and Reconstruction Estimation), a novel framework that extends the capabilities of pretrained multimodal models by introducing latent space reconstruction and uncertainty estimation for both reconstructed modalities and downstream tasks. Our method is architecture-agnostic, reconstructs missing modalities, and delivers reliable uncertainty estimates, improving both interpretability and performance. SURE introduces a unique Pearson Correlation-based loss and applies statistical error propagation in deep networks for the first time, allowing precise quantification of uncertainties from missing data and model predictions. Extensive experiments across tasks such as sentiment analysis, genre classification, and action recognition show that SURE consistently achieves state-of-the-art performance, ensuring robust predictions even in the presence of incomplete data.

Leveraging Perfect Multimodal Alignment and Gaussian Assumptions for Cross-modal Transfer

Multimodal alignment aims to construct a joint latent vector space where two modalities representing the same concept map to the same vector. We formulate this as an inverse problem and show that under certain conditions perfect alignment can be achieved. We then address a specific application of alignment referred to as cross-modal transfer. Unsupervised cross-modal transfer aims to leverage a model trained with one modality to perform inference on another modality, without any labeled fine-tuning on the new modality. Assuming that semantic classes are represented as a mixture of Gaussians in the latent space, we show how cross-modal transfer can be performed by projecting the data points from the representation space onto different subspaces representing each modality. Our experiments on synthetic multimodal Gaussian data verify the effectiveness of our perfect alignment and cross-modal transfer method. We hope these findings inspire further exploration of the applications of perfect alignment and the use of Gaussian models for cross-modal learning.

GIST: Towards Photorealistic Style Transfer via Multiscale Geometric Representations

State-of-the-art Style Transfer methods often leverage pre-trained encoders optimized for discriminative tasks, which may not be ideal for image synthesis. This can result in significant artifacts and loss of photorealism. Motivated by the ability of multiscale geometric image representations to capture fine-grained details and global structure, we propose GIST: Geometric-based Image Style Transfer, a novel Style Transfer technique that exploits the geometric properties of content and style images. GIST replaces the standard Neural Style Transfer autoencoding framework with a multiscale image expansion, preserving scene details without the need for post-processing or training. Our method matches multiresolution and multidirectional representations such as Wavelets and Contourlets by solving an optimal transport problem, leading to an efficient texture transferring. Experiments show that GIST is on-par or outperforms recent photorealistic Style Transfer approaches while significantly reducing the processing time with no model training.

R.I.P.: A Simple Black-box Attack on Continual Test-time Adaptation

Test-time adaptation (TTA) has emerged as a promising solution to tackle the continual domain shift in machine learning by allowing model parameters to change at test time, via self-supervised learning on unlabeled testing data. At the same time, it unfortunately opens the door to unforeseen vulnerabilities for degradation over time. Through a simple theoretical continual TTA model, we successfully identify a risk in the sampling process of testing data that could easily degrade the performance of a continual TTA model. We name this risk as Reusing of Incorrect Prediction (RIP) that TTA attackers can employ or as a result of the unintended query from general TTA users. The risk posed by RIP is also highly realistic, as it does not require prior knowledge of model parameters or modification of testing samples. This simple requirement makes RIP as the first black-box TTA attack algorithm that stands out from existing white-box attempts. We extensively benchmark the performance of the most recent continual TTA approaches when facing the RIP attack, providing insights on its success, and laying out potential roadmaps that could enhance the resilience of future continual TTA systems.

Improving the Robustness of 3D Human Pose Estimation: A Benchmark and Learning from Noisy Input

Despite the promising performance of current 3D human pose estimation techniques, understanding and enhancing their generalization on challenging in-the-wild videos remain an open problem. In this work, we focus on the robustness of 2D-to-3D pose lifters. To this end, we develop two benchmark datasets, namely Human3.6M-C and HumanEva-I-C, to examine the robustness of video-based 3D pose lifters to a wide range of common video corruptions including temporary occlusion, motion blur, and pixel-level noise. We observe the poor generalization of state-of-the-art 3D pose lifters in the presence of corruption and establish two techniques to tackle this issue. First, we introduce Temporal Additive Gaussian Noise (TAGN) as a simple yet effective 2D input pose data augmentation. Additionally, to incorporate the confidence scores output by the 2D pose detectors, we design a confidence-aware convolution (CA-Conv) block. Extensively tested on corrupted videos, the proposed strategies consistently boost the robustness of 3D pose lifters and serve as new baselines for future research.

Persistent Test-time Adaptation in Episodic Testing Scenarios

Current test-time adaptation (TTA) approaches aim to adapt to environments that change continuously. Yet, when the environments not only change but also recur in a correlated manner over time, such as in the case of day-night surveillance cameras, it is unclear whether the adaptability of these methods is sustained after a long run. This study aims to examine the error accumulation of TTA models when they are repeatedly exposed to previous testing environments, proposing a novel testing setting called episodic TTA. To study this phenomenon, we design a simulation of TTA process on a simple yet representative $\epsilon$-perturbed Gaussian Mixture Model Classifier and derive the theoretical findings revealing the dataset- and algorithm-dependent factors that contribute to the gradual degeneration of TTA methods through time. Our investigation has led us to propose a method, named persistent TTA (PeTTA). PeTTA senses the model divergence towards a collapsing and adjusts the adaptation strategy of TTA, striking a balance between two primary objectives: adaptation and preventing model collapse. The stability of PeTTA in the face of episodic TTA scenarios has been demonstrated through a set of comprehensive experiments on various benchmarks.

Making Vision Transformers Truly Shift-Equivariant

For computer vision tasks, Vision Transformers (ViTs) have become one of the go-to deep net architectures. Despite being inspired by Convolutional Neural Networks (CNNs), ViTs remain sensitive to small shifts in the input image. To address this, we introduce novel designs for each of the modules in ViTs, such as tokenization, self-attention, patch merging, and positional encoding. With our proposed modules, we achieve truly shift-equivariant ViTs on four well-established models, namely, Swin, SwinV2, MViTv2, and CvT, both in theory and practice. Empirically, we tested these models on image classification and semantic segmentation, achieving competitive performance across three different datasets while maintaining 100% shift consistency.

MaskDiff: Modeling Mask Distribution with Diffusion Probabilistic Model for Few-Shot Instance Segmentation

Few-shot instance segmentation extends the few-shot learning paradigm to the instance segmentation task, which tries to segment instance objects from a query image with a few annotated examples of novel categories. Conventional approaches have attempted to address the task via prototype learning, known as point estimation. However, this mechanism is susceptible to noise and suffers from bias due to a significant scarcity of data. To overcome the disadvantages of the point estimation mechanism, we propose a novel approach, dubbed MaskDiff, which models the underlying conditional distribution of a binary mask, which is conditioned on an object region and $K$-shot information. Inspired by augmentation approaches that perturb data with Gaussian noise for populating low data density regions, we model the mask distribution with a diffusion probabilistic model. In addition, we propose to utilize classifier-free guided mask sampling to integrate category information into the binary mask generation process. Without bells and whistles, our proposed method consistently outperforms state-of-the-art methods on both base and novel classes of the COCO dataset while simultaneously being more stable than existing methods.

FedDCT: Federated Learning of Large Convolutional Neural Networks on Resource Constrained Devices using Divide and Co-Training

We introduce FedDCT, a novel distributed learning paradigm that enables the usage of large, high-performance CNNs on resource-limited edge devices. As opposed to traditional FL approaches, which require each client to train the full-size neural network independently during each training round, the proposed FedDCT allows a cluster of several clients to collaboratively train a large deep learning model by dividing it into an ensemble of several small sub-models and train them on multiple devices in parallel while maintaining privacy. In this co-training process, clients from the same cluster can also learn from each other, further improving their ensemble performance. In the aggregation stage, the server takes a weighted average of all the ensemble models trained by all the clusters. FedDCT reduces the memory requirements and allows low-end devices to participate in FL. We empirically conduct extensive experiments on standardized datasets, including CIFAR-10, CIFAR-100, and two real-world medical datasets HAM10000 and VAIPE. Experimental results show that FedDCT outperforms a set of current SOTA FL methods with interesting convergence behaviors. Furthermore, compared to other existing approaches, FedDCT achieves higher accuracy and substantially reduces the number of communication rounds (with $4-8$ times fewer memory requirements) to achieve the desired accuracy on the testing dataset without incurring any extra training cost on the server side.

Enhancing Few-shot Image Classification with Cosine Transformer

This paper addresses the few-shot image classification problem. One notable limitation of few-shot learning is the variation in describing the same category, which might result in a significant difference between small labeled support and large unlabeled query sets. Our approach is to obtain a relation heatmap between the two sets in order to label the latter one in a transductive setting manner. This can be solved by using cross-attention with the scaled dot-product mechanism. However, the magnitude differences between two separate sets of embedding vectors may cause a significant impact on the output attention map and affect model performance. We tackle this problem by improving the attention mechanism with cosine similarity. Specifically, we develop FS-CT (Few-shot Cosine Transformer), a few-shot image classification method based on prototypical embedding and transformer-based framework. The proposed Cosine attention improves FS-CT performances significantly from nearly 5% to over 20% in accuracy compared to the baseline scaled dot-product attention in various scenarios on three few-shot datasets mini-ImageNet, CUB-200, and CIFAR-FS. Additionally, we enhance the prototypical embedding for categorical representation with learnable weights before feeding them to the attention module. Our proposed method FS-CT along with the Cosine attention is simple to implement and can be applied for a wide range of applications. Our codes are available at https://github.com/vinuni-vishc/Few-Shot-Cosine-Transformer