Amir Rosenfeld

Amir Rosenfeld, Neta Glazer, Ethan Fetaya

Visual Language Models (VLMs) have achieved remarkable progress, yet their reliability under small, meaning-preserving input changes remains poorly understood. We present the first large-scale, systematic study of VLM robustness to benign visual and textual perturbations: pixel-level shifts, light geometric transformations, padded rescaling, paraphrasing, and multilingual rewrites that do not alter the underlying semantics of an image-question pair. Across a broad set of models and datasets, we find that modern VLMs are highly sensitive to such minor perturbations: a substantial fraction of samples change their predicted answer under at least one visual or textual modification. We characterize how this instability varies across perturbation types, question categories, and models, revealing that even state-of-the-art systems (e.g., GPT-4o, Gemini 2.0 Flash) frequently fail under shifts as small as a few pixels or harmless rephrasings. We further show that sample-level stability serves as a strong indicator of correctness: stable samples are consistently far more likely to be answered correctly. Leveraging this, we demonstrate that the stability patterns of small, accessible open-source models can be used to predict the correctness of much larger closed-source models with high precision. Our findings expose a fundamental fragility in current VLMs and highlight the need for robustness evaluations that go beyond adversarial perturbations, focusing instead on invariances that models should reliably uphold.

Idan Achituve, Hai Victor Habi, Amir Rosenfeld et al.

In image processing, solving inverse problems is the task of finding plausible reconstructions of an image that was corrupted by some (usually known) degradation operator. Commonly, this process is done using a generative image model that can guide the reconstruction towards solutions that appear natural. The success of diffusion models over the last few years has made them a leading candidate for this task. However, the sequential nature of diffusion models makes this conditional sampling process challenging. Furthermore, since diffusion models are often defined in the latent space of an autoencoder, the encoder-decoder transformations introduce additional difficulties. To address these challenges, we suggest a novel sampling method based on sequential Monte Carlo (SMC) in the latent space of diffusion models. We name our method LD-SMC. We define a generative model for the data using additional auxiliary observations and perform posterior inference with SMC sampling based on a reverse diffusion process. Empirical evaluations on ImageNet and FFHQ show the benefits of LD-SMC over competing methods in various inverse problem tasks and especially in challenging inpainting tasks.

Idit Diamant, Amir Rosenfeld, Idan Achituve et al.

Source-free domain adaptation aims to adapt a source-trained model to an unlabeled target domain without access to the source data. It has attracted growing attention in recent years, where existing approaches focus on self-training that usually includes pseudo-labeling techniques. In this paper, we introduce a novel noise-learning approach tailored to address noise distribution in domain adaptation settings and learn to de-confuse the pseudo-labels. More specifically, we learn a noise transition matrix of the pseudo-labels to capture the label corruption of each class and learn the underlying true label distribution. Estimating the noise transition matrix enables a better true class-posterior estimation, resulting in better prediction accuracy. We demonstrate the effectiveness of our approach when combined with several source-free domain adaptation methods: SHOT, SHOT++, and AaD. We obtain state-of-the-art results on three domain adaptation datasets: VisDA, DomainNet, and OfficeHome.

Itay Cohen, Ethan Fetaya, Amir Rosenfeld

Recent advances in computer vision have yielded models with strong performance on recognition benchmarks; however, significant gaps remain in comparison to human perception. One subtle ability is to judge whether an image looks like a given object without being an instance of that object. We study whether vision-language models such as CLIP capture this distinction. We curated a dataset named RoLA (Real or Lookalike) of real and lookalike exemplars (e.g., toys, statues, drawings, pareidolia) across multiple categories, and first evaluate a prompt-based baseline with paired "real"/"lookalike" prompts. We then estimate a direction in CLIP's embedding space that moves representations between real and lookalike. Applying this direction to image and text embeddings improves discrimination in cross-modal retrieval on Conceptual12M, and also enhances captions produced by a CLIP prefix captioner.

Jonathan S. Rosenfeld, Amir Rosenfeld, Yonatan Belinkov et al.

The dependency of the generalization error of neural networks on model and dataset size is of critical importance both in practice and for understanding the theory of neural networks. Nevertheless, the functional form of this dependency remains elusive. In this work, we present a functional form which approximates well the generalization error in practice. Capitalizing on the successful concept of model scaling (e.g., width, depth), we are able to simultaneously construct such a form and specify the exact models which can attain it across model/data scales. Our construction follows insights obtained from observations conducted over a range of model/data scales, in various model types and datasets, in vision and language tasks. We show that the form both fits the observations well across scales, and provides accurate predictions from small- to large-scale models and data.

Amir Rosenfeld, Richard Zemel, John K. Tsotsos

Predicting human perceptual similarity is a challenging subject of ongoing research. The visual process underlying this aspect of human vision is thought to employ multiple different levels of visual analysis (shapes, objects, texture, layout, color, etc). In this paper, we postulate that the perception of image similarity is not an explicitly learned capability, but rather one that is a byproduct of learning others. This claim is supported by leveraging representations learned from a diverse set of visual tasks and using them jointly to predict perceptual similarity. This is done via simple feature concatenation, without any further learning. Nevertheless, experiments performed on the challenging Totally-Looks-Like (TLL) benchmark significantly surpass recent baselines, closing much of the reported gap towards prediction of human perceptual similarity. We provide an analysis of these results and discuss them in a broader context of emergent visual capabilities and their implications on the course of machine-vision research.

Amir Rosenfeld, Richard Zemel, John K. Tsotsos

We showcase a family of common failures of state-of-the art object detectors. These are obtained by replacing image sub-regions by another sub-image that contains a trained object. We call this "object transplanting". Modifying an image in this manner is shown to have a non-local impact on object detection. Slight changes in object position can affect its identity according to an object detector as well as that of other objects in the image. We provide some analysis and suggest possible reasons for the reported phenomena.

Amir Rosenfeld, Markus D. Solbach, John K. Tsotsos

Perceptual judgment of image similarity by humans relies on rich internal representations ranging from low-level features to high-level concepts, scene properties and even cultural associations. However, existing methods and datasets attempting to explain perceived similarity use stimuli which arguably do not cover the full breadth of factors that affect human similarity judgments, even those geared toward this goal. We introduce a new dataset dubbed Totally-Looks-Like (TLL) after a popular entertainment website, which contains images paired by humans as being visually similar. The dataset contains 6016 image-pairs from the wild, shedding light upon a rich and diverse set of criteria employed by human beings. We conduct experiments to try to reproduce the pairings via features extracted from state-of-the-art deep convolutional neural networks, as well as additional human experiments to verify the consistency of the collected data. Though we create conditions to artificially make the matching task increasingly easier, we show that machine-extracted representations perform very poorly in terms of reproducing the matching selected by humans. We discuss and analyze these results, suggesting future directions for improvement of learned image representations.

Amir Rosenfeld, John K. Tsotsos

While great advances are made in pattern recognition and machine learning, the successes of such fields remain restricted to narrow applications and seem to break down when training data is scarce, a shift in domain occurs, or when intelligent reasoning is required for rapid adaptation to new environments. In this work, we list several of the shortcomings of modern machine-learning solutions, specifically in the contexts of computer vision and in reinforcement learning and suggest directions to explore in order to try to ameliorate these weaknesses.

Amir Rosenfeld, John K. Tsotsos

There is no denying the tremendous leap in the performance of machine learning methods in the past half-decade. Some might even say that specific sub-fields in pattern recognition, such as machine-vision, are as good as solved, reaching human and super-human levels. Arguably, lack of training data and computation power are all that stand between us and solving the remaining ones. In this position paper we underline cases in vision which are challenging to machines and even to human observers. This is to show limitations of contemporary models that are hard to ameliorate by following the current trend to increase training data, network capacity or computational power. Moreover, we claim that attempting to do so is in principle a suboptimal approach. We provide a taster of such examples in hope to encourage and challenge the machine learning community to develop new directions to solve the said difficulties.

Amir Rosenfeld, John K. Tsotsos

Training deep neural networks results in strong learned representations that show good generalization capabilities. In most cases, training involves iterative modification of all weights inside the network via back-propagation. In Extreme Learning Machines, it has been suggested to set the first layer of a network to fixed random values instead of learning it. In this paper, we propose to take this approach a step further and fix almost all layers of a deep convolutional neural network, allowing only a small portion of the weights to be learned. As our experiments show, fixing even the majority of the parameters of the network often results in performance which is on par with the performance of learning all of them. The implications of this intriguing property of deep neural networks are discussed and we suggest ways to harness it to create more robust representations.

Amir Rosenfeld, Mahdi Biparva, John K. Tsotsos

Visual priming is known to affect the human visual system to allow detection of scene elements, even those that may have been near unnoticeable before, such as the presence of camouflaged animals. This process has been shown to be an effect of top-down signaling in the visual system triggered by the said cue. In this paper, we propose a mechanism to mimic the process of priming in the context of object detection and segmentation. We view priming as having a modulatory, cue dependent effect on layers of features within a network. Our results show how such a process can be complementary to, and at times more effective than simple post-processing applied to the output of the network, notably so in cases where the object is hard to detect such as in severe noise. Moreover, we find the effects of priming are sometimes stronger when early visual layers are affected. Overall, our experiments confirm that top-down signals can go a long way in improving object detection and segmentation.

Amir Rosenfeld, John K. Tsotsos

Given an existing trained neural network, it is often desirable to learn new capabilities without hindering performance of those already learned. Existing approaches either learn sub-optimal solutions, require joint training, or incur a substantial increment in the number of parameters for each added domain, typically as many as the original network. We propose a method called \emph{Deep Adaptation Networks} (DAN) that constrains newly learned filters to be linear combinations of existing ones. DANs precisely preserve performance on the original domain, require a fraction (typically 13\%, dependent on network architecture) of the number of parameters compared to standard fine-tuning procedures and converge in less cycles of training to a comparable or better level of performance. When coupled with standard network quantization techniques, we further reduce the parameter cost to around 3\% of the original with negligible or no loss in accuracy. The learned architecture can be controlled to switch between various learned representations, enabling a single network to solve a task from multiple different domains. We conduct extensive experiments showing the effectiveness of our method on a range of image classification tasks and explore different aspects of its behavior.

Amir Rosenfeld, Shimon Ullman

Classes in natural images tend to follow long tail distributions. This is problematic when there are insufficient training examples for rare classes. This effect is emphasized in compound classes, involving the conjunction of several concepts, such as those appearing in action-recognition datasets. In this paper, we propose to address this issue by learning how to utilize common visual concepts which are readily available. We detect the presence of prominent concepts in images and use them to infer the target labels instead of using visual features directly, combining tools from vision and natural-language processing. We validate our method on the recently introduced HICO dataset reaching a mAP of 31.54\% and on the Stanford-40 Actions dataset, where the proposed method outperforms that obtained by direct visual features, obtaining an accuracy 83.12\%. Moreover, the method provides for each class a semantically meaningful list of keywords and relevant image regions relating it to its constituent concepts.

Amir Rosenfeld, Shimon Ullman

Convolutional neural networks have been shown to develop internal representations, which correspond closely to semantically meaningful objects and parts, although trained solely on class labels. Class Activation Mapping (CAM) is a recent method that makes it possible to easily highlight the image regions contributing to a network's classification decision. We build upon these two developments to enable a network to re-examine informative image regions, which we term introspection. We propose a weakly-supervised iterative scheme, which shifts its center of attention to increasingly discriminative regions as it progresses, by alternating stages of classification and introspection. We evaluate our method and show its effectiveness over a range of several datasets, where we obtain competitive or state-of-the-art results: on Stanford-40 Actions, we set a new state-of the art of 81.74%. On FGVC-Aircraft and the Stanford Dogs dataset, we show consistent improvements over baselines, some of which include significantly more supervision.

Amir Rosenfeld, Shimon Ullman

Action recognition in still images has seen major improvement in recent years due to advances in human pose estimation, object recognition and stronger feature representations. However, there are still many cases in which performance remains far from that of humans. In this paper, we approach the problem by learning explicitly, and then integrating three components of transitive actions: (1) the human body part relevant to the action (2) the object being acted upon and (3) the specific form of interaction between the person and the object. The process uses class-specific features and relations not used in the past for action recognition and which use inherently two cycles in the process unlike most standard approaches. We focus on face-related actions (FRA), a subset of actions that includes several currently challenging categories. We present an average relative improvement of 52% over state-of-the art. We also make a new benchmark publicly available.

Amir Rosenfeld, Shimon Ullman

Action recognition in still images has seen major improvement in recent years due to advances in human pose estimation, object recognition and stronger feature representations produced by deep neural networks. However, there are still many cases in which performance remains far from that of humans. A major difficulty arises in distinguishing between transitive actions in which the overall actor pose is similar, and recognition therefore depends on details of the grasp and the object, which may be largely occluded. In this paper we demonstrate how recognition is improved by obtaining precise localization of the action-object and consequently extracting details of the object shape together with the actor-object interaction. To obtain exact localization of the action object and its interaction with the actor, we employ a coarse-to-fine approach which combines semantic segmentation and contextual features, in successive stages. We focus on (but are not limited) to face-related actions, a set of actions that includes several currently challenging categories. We present an average relative improvement of 35% over state-of-the art and validate through experimentation the effectiveness of our approach.