BayesAdapter: enhanced uncertainty estimation in CLIP few-shot adaptation

The emergence of large pre-trained vision-language models (VLMs) represents a paradigm shift in machine learning, with unprecedented results in a broad span of visual recognition tasks. CLIP, one of the most popular VLMs, has exhibited remarkable zero-shot and transfer learning capabilities in classification. To transfer CLIP to downstream tasks, adapters constitute a parameter-efficient approach that avoids backpropagation through the large model (unlike related prompt learning methods). However, CLIP adapters have been developed to target discriminative performance, and the quality of their uncertainty estimates has been overlooked. In this work we show that the discriminative performance of state-of-the-art CLIP adapters does not always correlate with their uncertainty estimation capabilities, which are essential for a safe deployment in real-world scenarios. We also demonstrate that one of such adapters is obtained through MAP inference from a more general probabilistic framework. Based on this observation we introduce BayesAdapter, which leverages Bayesian inference to estimate a full probability distribution instead of a single point, better capturing the variability inherent in the parameter space. In a comprehensive empirical evaluation we show that our approach obtains high quality uncertainty estimates in the predictions, standing out in calibration and selective classification. Our code is publicly available at: https://github.com/pablomorales92/BayesAdapter.

Are foundation models for computer vision good conformal predictors?

Recent advances in self-supervision and constrastive learning have brought the performance of foundation models to unprecedented levels in a variety of tasks. Fueled by this progress, these models are becoming the prevailing approach for a wide array of real-world vision problems, including risk-sensitive and high-stakes applications. However, ensuring safe deployment in these scenarios requires a more comprehensive understanding of their uncertainty modeling capabilities, which has been barely explored. In this work, we delve into the behavior of vision and vision-language foundation models under Conformal Prediction (CP), a statistical framework that provides theoretical guarantees of marginal coverage of the true class. Across extensive experiments including popular vision classification benchmarks, well-known foundation vision models, and three CP methods, our findings reveal that foundation models are well-suited for conformalization procedures, particularly those integrating Vision Transformers. Furthermore, we show that calibrating the confidence predictions of these models leads to efficiency degradation of the conformal set on adaptive CP methods. In contrast, few-shot adaptation to downstream tasks generally enhances conformal scores, where we identify Adapters as a better conformable alternative compared to Prompt Learning strategies. Our empirical study identifies APS as particularly promising in the context of vision foundation models, as it does not violate the marginal coverage property across multiple challenging, yet realistic scenarios.

Words Matter: Leveraging Individual Text Embeddings for Code Generation in CLIP Test-Time Adaptation

Vision-language foundation models, such as CLIP, have shown unprecedented zero-shot performance across a wide range of tasks. Nevertheless, these models may be unreliable under distributional shifts, as their performance is significantly degraded. In this work, we explore how to efficiently leverage class text information to mitigate these distribution drifts encountered by large pre-trained vision-language models (VLMs) during test-time inference. In particular, we propose to generate pseudo-labels for the test-time samples by exploiting generic class text embeddings as fixed centroids of a label assignment problem, which is efficiently solved with Optimal Transport. Furthermore, the proposed adaptation method (CLIP-OT) integrates a multiple template knowledge distillation approach, which replicates multi-view contrastive learning strategies in unsupervised representation learning but without incurring additional computational complexity. Extensive experiments on multiple popular test-time adaptation benchmarks presenting diverse complexity empirically show the superiority of CLIP-OT, achieving performance gains of up to 7% over recent state-of-the-art methods, yet being computationally and memory efficient.

ReC-TTT: Contrastive Feature Reconstruction for Test-Time Training

The remarkable progress in deep learning (DL) showcases outstanding results in various computer vision tasks. However, adaptation to real-time variations in data distributions remains an important challenge. Test-Time Training (TTT) was proposed as an effective solution to this issue, which increases the generalization ability of trained models by adding an auxiliary task at train time and then using its loss at test time to adapt the model. Inspired by the recent achievements of contrastive representation learning in unsupervised tasks, we propose ReC-TTT, a test-time training technique that can adapt a DL model to new unseen domains by generating discriminative views of the input data. ReC-TTT uses cross-reconstruction as an auxiliary task between a frozen encoder and two trainable encoders, taking advantage of a single shared decoder. This enables, at test time, to adapt the encoders to extract features that will be correctly reconstructed by the decoder that, in this phase, is frozen on the source domain. Experimental results show that ReC-TTT achieves better results than other state-of-the-art techniques in most domain shift classification challenges.

Few-shot Adaptation of Medical Vision-Language Models

Integrating image and text data through multi-modal learning has emerged as a new approach in medical imaging research, following its successful deployment in computer vision. While considerable efforts have been dedicated to establishing medical foundation models and their zero-shot transfer to downstream tasks, the popular few-shot setting remains relatively unexplored. Following on from the currently strong emergence of this setting in computer vision, we introduce the first structured benchmark for adapting medical vision-language models (VLMs) in a strict few-shot regime and investigate various adaptation strategies commonly used in the context of natural images. Furthermore, we evaluate a simple generalization of the linear-probe adaptation baseline, which seeks an optimal blending of the visual prototypes and text embeddings via learnable class-wise multipliers. Surprisingly, such a text-informed linear probe yields competitive performances in comparison to convoluted prompt-learning and adapter-based strategies, while running considerably faster and accommodating the black-box setting. Our extensive experiments span three different medical modalities and specialized foundation models, nine downstream tasks, and several state-of-the-art few-shot adaptation methods. We made our benchmark and code publicly available to trigger further developments in this emergent subject: \url{https://github.com/FereshteShakeri/few-shot-MedVLMs}.

Harmonizing Flows: Leveraging normalizing flows for unsupervised and source-free MRI harmonization

Lack of standardization and various intrinsic parameters for magnetic resonance (MR) image acquisition results in heterogeneous images across different sites and devices, which adversely affects the generalization of deep neural networks. To alleviate this issue, this work proposes a novel unsupervised harmonization framework that leverages normalizing flows to align MR images, thereby emulating the distribution of a source domain. The proposed strategy comprises three key steps. Initially, a normalizing flow network is trained to capture the distribution characteristics of the source domain. Then, we train a shallow harmonizer network to reconstruct images from the source domain via their augmented counterparts. Finally, during inference, the harmonizer network is updated to ensure that the output images conform to the learned source domain distribution, as modeled by the normalizing flow network. Our approach, which is unsupervised, source-free, and task-agnostic is assessed in the context of both adults and neonatal cross-domain brain MRI segmentation, as well as neonatal brain age estimation, demonstrating its generalizability across tasks and population demographics. The results underscore its superior performance compared to existing methodologies. The code is available at https://github.com/farzad-bz/Harmonizing-Flows

Robust Calibration of Large Vision-Language Adapters

This paper addresses the critical issue of miscalibration in CLIP-based model adaptation, particularly in the challenging scenario of out-of-distribution (OOD) samples, which has been overlooked in the existing literature on CLIP adaptation. We empirically demonstrate that popular CLIP adaptation approaches, such as Adapters, Prompt Learning, and Test-Time Adaptation, substantially degrade the calibration capabilities of the zero-shot baseline in the presence of distributional drift. We identify the increase in logit ranges as the underlying cause of miscalibration of CLIP adaptation methods, contrasting with previous work on calibrating fully-supervised models. Motivated by these observations, we present a simple and model-agnostic solution to mitigate miscalibration, by scaling the logit range of each sample to its zero-shot prediction logits. We explore three different alternatives to achieve this, which can be either integrated during adaptation or directly used at inference time. Comprehensive experiments on popular OOD classification benchmarks demonstrate the effectiveness of the proposed approaches in mitigating miscalibration while maintaining discriminative performance, whose improvements are consistent across the three families of these increasingly popular approaches. The code is publicly available at: https://github.com/Bala93/CLIPCalib

Pay Attention to Your Neighbours: Training-Free Open-Vocabulary Semantic Segmentation

Despite the significant progress in deep learning for dense visual recognition problems, such as semantic segmentation, traditional methods are constrained by fixed class sets. Meanwhile, vision-language foundation models, such as CLIP, have showcased remarkable effectiveness in numerous zero-shot image-level tasks, owing to their robust generalizability. Recently, a body of work has investigated utilizing these models in open-vocabulary semantic segmentation (OVSS). However, existing approaches often rely on impractical supervised pre-training or access to additional pre-trained networks. In this work, we propose a strong baseline for training-free OVSS, termed Neighbour-Aware CLIP (NACLIP), representing a straightforward adaptation of CLIP tailored for this scenario. Our method enforces localization of patches in the self-attention of CLIP's vision transformer which, despite being crucial for dense prediction tasks, has been overlooked in the OVSS literature. By incorporating design choices favouring segmentation, our approach significantly improves performance without requiring additional data, auxiliary pre-trained networks, or extensive hyperparameter tuning, making it highly practical for real-world applications. Experiments are performed on 8 popular semantic segmentation benchmarks, yielding state-of-the-art performance on most scenarios. Our code is publicly available at https://github.com/sinahmr/NACLIP .

LP++: A Surprisingly Strong Linear Probe for Few-Shot CLIP

In a recent, strongly emergent literature on few-shot CLIP adaptation, Linear Probe (LP) has been often reported as a weak baseline. This has motivated intensive research building convoluted prompt learning or feature adaptation strategies. In this work, we propose and examine from convex-optimization perspectives a generalization of the standard LP baseline, in which the linear classifier weights are learnable functions of the text embedding, with class-wise multipliers blending image and text knowledge. As our objective function depends on two types of variables, i.e., the class visual prototypes and the learnable blending parameters, we propose a computationally efficient block coordinate Majorize-Minimize (MM) descent algorithm. In our full-batch MM optimizer, which we coin LP++, step sizes are implicit, unlike standard gradient descent practices where learning rates are intensively searched over validation sets. By examining the mathematical properties of our loss (e.g., Lipschitz gradient continuity), we build majorizing functions yielding data-driven learning rates and derive approximations of the loss's minima, which provide data-informed initialization of the variables. Our image-language objective function, along with these non-trivial optimization insights and ingredients, yields, surprisingly, highly competitive few-shot CLIP performances. Furthermore, LP++ operates in black-box, relaxes intensive validation searches for the optimization hyper-parameters, and runs orders-of-magnitudes faster than state-of-the-art few-shot CLIP adaptation methods. Our code is available at: \url{https://github.com/FereshteShakeri/FewShot-CLIP-Strong-Baseline.git}.

Do not trust what you trust: Miscalibration in Semi-supervised Learning

State-of-the-art semi-supervised learning (SSL) approaches rely on highly confident predictions to serve as pseudo-labels that guide the training on unlabeled samples. An inherent drawback of this strategy stems from the quality of the uncertainty estimates, as pseudo-labels are filtered only based on their degree of uncertainty, regardless of the correctness of their predictions. Thus, assessing and enhancing the uncertainty of network predictions is of paramount importance in the pseudo-labeling process. In this work, we empirically demonstrate that SSL methods based on pseudo-labels are significantly miscalibrated, and formally demonstrate the minimization of the min-entropy, a lower bound of the Shannon entropy, as a potential cause for miscalibration. To alleviate this issue, we integrate a simple penalty term, which enforces the logit distances of the predictions on unlabeled samples to remain low, preventing the network predictions to become overconfident. Comprehensive experiments on a variety of SSL image classification benchmarks demonstrate that the proposed solution systematically improves the calibration performance of relevant SSL models, while also enhancing their discriminative power, being an appealing addition to tackle SSL tasks.