Navve Wasserman, Roi Pony, Oshri Naparstek, Adi Raz Goldfarb, Eli Schwartz, Udi Barzelay, Leonid Karlinsky. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

We introduce XTRA, a vision model pre-trained with a novel auto-regressive objective that significantly enhances both sample and parameter efficiency compared to previous auto-regressive image models. Unlike contrastive or masked image modeling methods, which have not been demonstrated as having consistent scaling behavior on unbalanced internet data, auto-regressive vision models exhibit scalable and promising performance as model and dataset size increase. In contrast to standard auto-regressive models, XTRA employs a Block Causal Mask, where each Block represents k × k tokens rather than relying on a standard causal mask. By reconstructing pixel values block by block, XTRA captures higher-level structural patterns over larger image regions. Predicting on blocks allows the model to learn relationships across broader areas of pixels, enabling more abstract and semantically meaningful representations than traditional next-token prediction. This simple modification yields two key results. First, XTRA is sample- efficient. Despite being trained on 152 fewer samples (13.1M vs. 2B), XTRA ViT-H/14 surpasses the×top-1 average accuracy of the previous state-of-the-art auto-regressive model across 15 diverse image recognition benchmarks. Second, XTRA is parameter-efficient. Compared to auto-regressive models trained on ImageNet-1k, XTRA ViT-B/16 outperforms in linear and attentive probing tasks, using 7–16× fewer parameters (85M vs. 1.36B/0.63B).

Recent advances in audio-visual learning have shown promising results in learning representations across modalities. However, most approaches rely on global audio representations that fail to capture fine-grained temporal correspondences with visual frames. Additionally, existing methods often struggle with conflicting optimization objectives when trying to jointly learn reconstruction and cross-modal alignment. In this work, we propose CAV-MAE Sync as a simple yet effective extension of the original CAV-MAE [14] framework for self-supervised audio-visual learning. We address three key challenges: First, we tackle the granularity mismatch between modalities by treating audio as a temporal sequence aligned with video frames, rather than using global representations. Second, we resolve conflicting optimization goals by separating contrastive and reconstruction objectives through dedicated global tokens. Third, we improve spatial localization by introducing learnable register tokens that reduce the semantic load on patch tokens. We evaluate the proposed approach on AudioSet, VGG Sound, and the ADE20K Sound dataset on zero-shot retrieval, classification, and localization tasks demonstrating state-of-the-art performance and outperforming more complex architectures. Code is available at https://github.com/edsonroteia/cav-mae-sync.

Vision-Language Models (VLMs) have shown remarkable capabilities across diverse visual tasks, including image recognition, video understanding, and Visual Question Answering (VQA) when explicitly trained for these tasks. Despite these advances, we find that present-day VLMs (including the proprietary GPT-4o) lack a fundamental cognitive ability: learning to localize specific objects in a scene by taking into account the context. In this work, we focus on the task of few-shot personalized localization, where a model is given a small set of annotated images (in-context examples) -- each with a category label and bounding box -- and is tasked with localizing the same object type in a query image. Personalized localization can be particularly important in cases of ambiguity of several related objects that can respond to a text or an object that is hard to describe with words. To provoke personalized localization abilities in models, we present a data-centric solution that fine-tunes them using carefully curated data from video object tracking datasets. By leveraging sequences of frames tracking the same object across multiple shots, we simulate instruction-tuning dialogues that promote context awareness. To reinforce this, we introduce a novel regularization technique that replaces object labels with pseudo-names, ensuring the model relies on visual context rather than prior knowledge. Our method significantly enhances the few-shot localization performance of recent VLMs ranging from 7B to 72B in size, without sacrificing generalization, as demonstrated on several benchmarks tailored towards evaluating personalized localization abilities. This work is the first to explore and benchmark personalized few-shot localization for VLMs -- exposing critical weaknesses in present-day VLMs, and laying a foundation for future research in context-driven vision-language applications.

We present Self-MoE, an approach that transforms a monolithic LLM into a compositional, modular system of self-specialized experts, named MiXSE (MiXture of Self-specialized Experts). Our approach leverages self-specialization, which constructs expert modules using self-generated synthetic data, each equipping a shared base LLM with distinct domain-specific capabilities, activated via self-optimized routing. This allows for dynamic and capability-specific handling of various target tasks, enhancing overall capabilities, without extensive human-labeled data and added parameters. Our empirical results reveal that specializing LLMs may exhibit potential trade-offs in performances on non-specialized tasks. On the other hand, our Self-MoE demonstrates substantial improvements (6.5%p on average) over the base LLM across diverse benchmarks such as knowledge, reasoning, math, and coding. It also consistently outperforms other methods, including instance merging and weight merging, while offering better flexibility and interpretability by design with semantic experts and routing. Our findings highlight the critical role of modularity, the applicability of Self-MoE to multiple base LLMs, and the potential of self-improvement in achieving efficient, scalable, and adaptable systems.

The large-scale training of multi-modal models on data scraped from the web has shown outstanding utility in infusing these models with the required world knowledge to perform effectively on multiple downstream tasks. However, one downside of scraping data from the web can be the potential sacrifice of the benchmarks on which the abilities of these models are often evaluated. To safeguard against test data contamination and to truly test the abilities of these foundation models we propose LiveXiv: A scalable evolving live benchmark based on scientific ArXiv papers. LiveXiv accesses domain-specific manuscripts at any given timestamp and proposes to automatically generate visual question-answer pairs (VQA). This is done without any human-in-the-loop, using the multi-modal content in the manuscripts, like graphs, charts, and tables. Moreover, we introduce an efficient evaluation approach that estimates the performance of all models on the evolving benchmark using evaluations of only a subset of models. This significantly reduces the overall evaluation cost. We benchmark multiple open and proprietary Large Multi-modal Models (LMMs) on the first version of our benchmark, showing its challenging nature and exposing the models’ true abilities, avoiding contamination. Lastly, in our commitment to high quality, we have collected and evaluated a manually verified subset. By comparing its overall results to our automatic annotations, we have found that the performance variance is indeed minimal (<2.5%). Our dataset is available online on HuggingFace, and our code is available here.

Vision-Language models (VLMs) have proven to be effective at aligning image and text representations, producing superior zero-shot results when transferred to many downstream tasks. However, these representations suffer from some key shortcomings in understanding Compositional Language Concepts (CLC), such as recognizing objects' attributes, states, and relations between different objects. Moreover, VLMs typically have poor interpretability, making it challenging to debug and mitigate compositional-understanding failures. In this work, we introduce the architecture and training technique of Tree-augmented Vision-Language (3VL) model accompanied by our proposed Anchor inference method and Differential Relevance (DiRe) interpretability tool. By expanding the text of an arbitrary image-text pair into a hierarchical tree structure using language analysis tools, 3VL allows the induction of this structure into the visual representation learned by the model, enhancing its interpretability and compositional reasoning. Additionally, we show how Anchor, a simple technique for text unification, can be used to filter nuisance factors while increasing CLC understanding performance, e.g., on the fundamental VL-Checklist benchmark. We also show how DiRe, which performs a differential comparison between VLM relevancy maps, enables us to generate compelling visualizations of the reasons for a model's success or failure.

Language models struggle with handling numerical data and performing arithmetic operations.We hypothesize that this limitation can be partially attributed to non-intuitive textual numbers representation.When a digit is read or generated by a causal language model it does not know its place value (e.g.thousands vs. hundreds) until the entire number is processed.To address this issue, we propose a simple adjustment to how numbers are represented by including the count of digits before each number.For instance, instead of "42", we suggest using "2:42" as the new format.This approach, which we term NumeroLogic, offers an added advantage in number generation by serving as a Chain of Thought (CoT).By requiring the model to consider the number of digits first, it enhances the reasoning process before generating the actual number.We use arithmetic tasks to demonstrate the effectiveness of the NumeroLogic formatting.We further demonstrate NumeroLogic applicability to general natural language modeling, improving language understanding performance in the MMLU benchmark.

The ability of artificial intelligence (AI) systems to perceive and comprehend audio signals is crucial for many applications. Although significant progress has been made in this area since the development of AudioSet, most existing models are designed to map audio inputs to pre-defined, discrete sound label sets. In contrast, humans possess the ability to not only classify sounds into general categories, but also to listen to the finer details of the sounds, explain the reason for the predictions, think about what the sound infers, and understand the scene and what action needs to be taken, if any. Such capabilities beyond perception are not yet present in existing audio models. On the other hand, modern large language models (LLMs) exhibit emerging reasoning ability but they lack audio perception capabilities. Therefore, we ask the question: can we build a model that has both audio perception and reasoning ability? In this paper, we propose a new audio foundation model, called LTU (Listen, Think, and Understand). To train LTU, we created a new OpenAQA-5M dataset consisting of 1.9 million closed-ended and 3.7 million open-ended, diverse (audio, question, answer) tuples, and have used an autoregressive training framework with a perception-to-understanding curriculum. LTU demonstrates strong performance and generalization ability on conventional audio tasks such as classification and captioning. More importantly, it exhibits emerging audio reasoning and comprehension abilities that are absent in existing audio models. To the best of our knowledge, LTU is the first multimodal large language model that focuses on general audio (rather than just speech) understanding.

We address the problem of scaling up the production of media content, including commentary and personalized news stories, for large-scale sports and music events worldwide. Our approach relies on generative AI models to transform a large volume of multimodal data (e.g., videos, articles, real-time scoring feeds, statistics, and fact sheets) into coherent and fluent text. Based on this approach, we introduce, for the first time, an AI commentary system, which was deployed to produce automated narrations for highlight packages at the 2023 US Open, Wimbledon, and Masters tournaments. In the same vein, our solution was extended to create personalized content for ESPN Fantasy Football and stories about music artists for the GRAMMY awards. These applications were built using a common software architecture achieved a 15x speed improvement with an average Rouge-L of 82.00 and perplexity of 6.6. Our work was successfully deployed at the aforementioned events, supporting 90 million fans around the world with 8 billion page views, continuously pushing the bounds on what is possible at the intersection of sports, entertainment, and AI.

Compositional Reasoning (CR) entails grasping the significance of attributes, relations, and word order. Recent Vision-Language Models (VLMs), comprising a visual encoder and a Large Language Model (LLM) decoder, have demonstrated remarkable proficiency in such reasoning tasks. This prompts a crucial question: have VLMs effectively tackled the CR challenge? We conjecture that existing CR benchmarks may not adequately push the boundaries of modern VLMs due to the reliance on an LLM only negative text generation pipeline. Consequently, the negatives produced either appear as outliers from the natural language distribution learned by VLMs' LLM decoders or as improbable within the corresponding image context. To address these limitations, we introduce ConMe\footnote{ConMe is an abbreviation for Confuse Me.} -- a compositional reasoning benchmark and a novel data generation pipeline leveraging VLMs to produce `hard CR Q&A'. Through a new concept of VLMs conversing with each other to collaboratively expose their weaknesses, our pipeline autonomously generates, evaluates, and selects challenging compositional reasoning questions, establishing a robust CR benchmark, also subsequently validated manually. Our benchmark provokes a noteworthy, up to 33%, decrease in CR performance compared to preceding benchmarks, reinstating the CR challenge even for state-of-the-art VLMs.

The recent success of interleaved Large Multimodal Models (LMMs) in few-shot learning suggests that in-context learning (ICL) with many examples can be promising for learning new tasks. However, this many-shot multimodal ICL setting has one crucial problem: it is fundamentally limited by the model's context length set at pretraining. The problem is especially prominent in the multimodal domain, which processes both text and images, requiring additional tokens. This motivates the need for a multimodal method to compress many shots into fewer tokens without finetuning. In this work, we enable LMMs to perform multimodal, many-shot in-context learning by leveraging Multimodal Task Vectors (MTV)---compact implicit representations of in-context examples compressed in the model's attention heads. Specifically, we first demonstrate the existence of such MTV in LMMs and then leverage these extracted MTV to enable many-shot in-context learning for various vision-and-language tasks. Our experiments suggest that MTV can scale in performance with the number of compressed shots and generalize to similar out-of-domain tasks without additional context length for inference. Code: https://github.com/Brandon3964/MultiModal-Task-Vector

State-space models (SSMs) have emerged as an alternative to Transformers for audio modeling due to their high computational efficiency with long inputs. While recent efforts on Audio SSMs have reported encouraging results, two main limitations remain: First, in 10 -second short audio tagging tasks, Audio SSMs still underperform compared to Transformer-based models such as Audio Spectrogram Transformer (AST). Second, although Audio SSMs theoretically support long audio inputs, their actual performance with long audio has not been thoroughly evaluated. To address these limitations, in this paper, 1) We applied knowledge distillation in audio space model training, resulting in a model called Knowledge Distilled Audio SSM (DASS). To the best of our knowledge, it is the first SSM that outperforms the Transformers on AudioSet and achieves an mAP of 47.6; and 2) We designed a new test called Audio Needle In A Haystack (Audio NIAH). We find that DASS, trained with only 10 -second audio clips, can retrieve sound events in audio recordings up to 2.5 hours long, while the AST model fails when the input is just 50 seconds, demonstrating state-space models are indeed more duration scalable.

Action recognition models have achieved impressive results by incorporating scene-level annotations, such as objects, their relations, 3D structure, and more. However, obtaining annotations of scene structure for videos requires a significant amount of effort to gather and annotate, making these methods expensive to train. In contrast, synthetic datasets generated by graphics engines provide powerful alternatives for generating scene-level annotations across multiple tasks. In this work, we propose an approach to leverage synthetic scene data for improving video understanding. We present a multi-task prompt learning approach for video transformers, where a shared video transformer backbone is enhanced by a small set of specialized parameters for each task. Specifically, we add a set of "task prompts", each corresponding to a different task, and let each prompt predict task-related annotations. This design allows the model to capture information shared among synthetic scene tasks as well as information shared between synthetic scene tasks and a real video downstream task throughout the entire network. We refer to this approach as "Promptonomy", since the prompts model task-related structure. We propose the PromptonomyViT model (PViT), a video transformer that incorporates various types of scene-level information from synthetic data using the "Promptonomy" approach. PViT shows strong performance improvements on multiple video understanding tasks and datasets. Project page: https://ofir1080.github.io/PromptonomyViT

Humans are surrounded by audio signals that include both speech and non-speech sounds. The recognition and understanding of speech and non-speech audio events, along with a profound comprehension of the relationship between them, constitute fundamental cognitive capabilities. For the first time, we build a machine learning model, called LTU-AS, that has a conceptually similar universal audio perception and advanced reasoning ability. Specifically, by integrating Whisper [1] as a perception module and LLaMA [2] as a reasoning module, LTU-AS can simultaneously recognize and jointly understand spoken text, speech paralinguistics, and non-speech audio events - almost everything perceivable from audio signals.

Vision and Language ($VL$) models have demonstrated remarkable zero-shot performance in a variety of tasks. However, some aspects of complex language understanding still remain a challenge. We introduce the collective notion of Structured Vision & Language Concepts (SVLC) which includes object attributes, relations, and states which are present in the text and visible in the image. Recent studies have shown that even the best$VL$models struggle with SVLC. A possible way of fixing this issue is by collecting dedicated datasets for teaching each SVLC type, yet this might be expensive and time-consuming. Instead, we propose a more elegant data-driven approach for enhancing$VL$models' understanding of SVLCs that makes more effective use of existing$VL$pre-training datasets and does not require any additional data. While automatic understanding of image structure still remains largely unsolved, language structure is much better modeled and understood, allowing for its effective utilization in teaching$VL$models. In this paper, we propose various techniques based on language structure understanding that can be used to manipulate the textual part of off-the-shelf paired$VL$datasets.$VL$models trained with the updated data exhibit a significant improvement of up to 15% in their SVLC understanding with only a mild degradation in their zero-shot capabilities both when training from scratch or fine-tuning a pre-trained model. Our code and pretrained models are available at: https://github.com/SivanDoveh/TSVLC

Recently, large-scale pre-trained Vision-and-Language (VL) foundation models have demonstrated remarkable capabilities in many zero-shot downstream tasks, achieving competitive results for recognizing objects defined by as little as short text prompts. However, it has also been shown that VL models are still brittle in Structured VL Concept (SVLC) reasoning, such as the ability to recognize object attributes, states, and inter-object relations. This leads to reasoning mistakes, which need to be corrected as they occur by teaching VL models the missing SVLC skills; often this must be done using private data where the issue was found, which naturally leads to a data-free continual (no task-id) VL learning setting. In this work, we introduce the first Continual Data-Free Structured VL Concepts Learning (ConStruct-VL) benchmark11Our code is publicly available at https://github.com/jamessealesmith/ConStruct-VL and show it is challenging for many existing data-free CL strategies. We, therefore, propose a data-free method comprised of a new approach of Adversarial Pseudo-Replay (APR) which generates adversarial reminders of past tasks from past task models. To use this method efficiently, we also propose a continual parameter-efficient Layered-LoRA (LaLo) neural architecture allowing no-memory-cost access to all past models at train time. We show this approach outperforms all data-free methods by as much as ~ 7% while even matching some levels of experience-replay (prohibitive for applications where data-privacy must be preserved).

Computer vision models suffer from a phenomenon known as catastrophic forgetting when learning novel concepts from continuously shifting training data. Typical solutions for this continual learning problem require extensive rehearsal of previously seen data, which increases memory costs and may violate data privacy. Recently, the emergence of large-scale pre-trained vision transformer models has enabled prompting approaches as an alternative to data-rehearsal. These approaches rely on a key-query mechanism to generate prompts and have been found to be highly resistant to catastrophic forgetting in the well-established rehearsal-free continual learning setting. However, the key mechanism of these methods is not trained end-to-end with the task sequence. Our experiments show that this leads to a reduction in their plasticity, hence sacrificing new task accuracy, and inability to benefit from expanded parameter capacity. We instead propose to learn a set of prompt components which are assembled with input-conditioned weights to produce input-conditioned prompts, resulting in a novel attention-based end-to-end key-query scheme. Our experiments show that we outperform the current SOTA method DualPrompt on established benchmarks by as much as 4.5% in average final accuracy. We also outperform the state of art by as much as 4.4% accuracy on a continual learning benchmark which contains both class-incremental and domain-incremental task shifts, corresponding to many practical settings. Our code is available at https://github.com/GT-RIPL/CODA-Prompt

Roei Herzig, Alon Mendelson, Leonid Karlinsky, Assaf Arbelle, Rogerio Feris, Trevor Darrell, Amir Globerson. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

Multilingual text-video retrieval methods have improved significantly in recent years, but the performance for languages other than English still lags. We propose a Cross-Lingual Cross-Modal Knowledge Distillation method to improve multilingual text-video retrieval. Inspired by the fact that English text-video retrieval outperforms other languages, we train a student model using input text in different languages to match the cross-modal predictions from teacher models using input text in English. We propose a cross entropy based objective which forces the distribution over the student’s text-video similarity scores to be similar to those of the teacher models. We introduce a new multilingual video dataset, Multi-YouCook2, by translating the English captions in the YouCook2 video dataset to 8 other languages. Our method improves multilingual text-video retrieval performance on Multi-YouCook2 and several other datasets such as Multi-MSRVTT and VATEX. We also conducted an analysis on the effectiveness of different multilingual text models as teachers.

Large scale Vision Language (VL) models have shown tremendous success in aligning representations between visual and text modalities. This enables remarkable progress in zero-shot recognition, image generation & editing, and many other exciting tasks. However, VL models tend to over-represent objects while paying much less attention to verbs, and require additional tuning on video data for best zero-shot action recognition performance. While previous work relied on large-scale, fully-annotated data, in this work we propose an unsupervised approach. We adapt a VL model for zero-shot and few-shot action recognition using a collection of unlabeled videos and an unpaired action dictionary. Based on that, we leverage Large Language Models and VL models to build a text bag for each unlabeled video via matching, text expansion and captioning. We use those bags in a Multiple Instance Learning setup to adapt an image-text backbone to video data. Although finetuned on unlabeled video data, our resulting models demonstrate high transferability to numerous unseen zero-shot downstream tasks, improving the base VL model performance by up to 14%, and even comparing favorably to fully-supervised baselines in both zero-shot and few-shot video recognition transfer. The code is released at https://github.com/wlin-at/MAXI.

Large-scale pre-trained Vision & Language (VL) models have shown remarkable performance in many applications, enabling replacing a fixed set of supported classes with zero-shot open vocabulary reasoning over (almost arbitrary) natural language prompts. However, recent works have uncovered a fundamental weakness of these models. For example, their difficulty to understand Visual Language Concepts (VLC) that go ‘beyond nouns’ such as the meaning of non-object words (e.g., attributes, actions, relations, states, etc.), or difficulty in performing compositional reasoning such as understanding the significance of the order of the words in a sentence. In this work, we investigate to which extent purely synthetic data could be leveraged to teach these models to overcome such shortcomings without compromising their zero-shot capabilities. We contribute Synthetic Visual Concepts (SyViC) - a million-scale synthetic dataset and data generation codebase allowing to generate additional suitable data to improve VLC understanding and compositional reasoning of VL models. Additionally, we propose a general VL finetuning strategy for effectively leveraging SyViC towards achieving these improvements. Our extensive experiments and ablations on VL-Checklist, Winoground, and ARO benchmarks demonstrate that it is possible to adapt strong pre-trained VL models with synthetic data significantly enhancing their VLC understanding (e.g. by 9.9% on ARO and 4.3% on VL-Checklist) with under 1% drop in their zero-shot accuracy.

In this paper, we focus on Whisper [1], a recent automatic speech recognition model trained with a massive 680k hour labeled speech corpus recorded in diverse conditions.We first show an interesting finding that while Whisper is very robust against real-world background sounds (e.g., music), its audio representation is actually not noise-invariant, but is instead highly correlated to non-speech sounds, indicating that Whisper recognizes speech conditioned on the noise type.With this finding, we build a unified audio tagging and speech recognition model Whisper-AT by freezing the backbone of Whisper, and training a lightweight audio tagging model on top of it.With <1% extra computational cost, Whisper-AT can recognize audio events, in addition to spoken text, in a single forward pass.

Vision and Language (VL) models offer an effective method for aligning representation spaces of images and text allowing for numerous applications such as cross-modal retrieval, visual and multi-hop question answering, captioning, and many more. However, the aligned image-text spaces learned by all the popular VL models are still suffering from the so-called 'object bias' - their representations behave as 'bags of nouns' mostly ignoring or downsizing the attributes, relations, and states of objects described/appearing in texts/images. Although some great attempts at fixing these `compositional reasoning' issues were proposed in the recent literature, the problem is still far from being solved. In this paper, we uncover two factors limiting the VL models' compositional reasoning performance. These two factors are properties of the paired VL dataset used for finetuning (or pre-training) the VL model: (i) the caption quality, or in other words 'image-alignment', of the texts; and (ii) the 'density' of the captions in the sense of mentioning all the details appearing on the image. We propose a fine-tuning approach for automatically treating these factors on a standard collection of paired VL data (CC3M). Applied to CLIP, we demonstrate its significant compositional reasoning performance increase of up to $\sim27$\% over the base model, up to $\sim20$\% over the strongest baseline, and by $6.7$\% on average. Our code is provided in the Supplementary and would be released upon acceptance.

Recently, large-scale pre-trained Vision and Language (VL) models have set a new state-of-the-art (SOTA) in zero-shot visual classification enabling open-vocabulary recognition of potentially unlimited set of categories defined as simple language prompts. However, despite these great advances, the performance of these zero-shot classifiers still falls short of the results of dedicated (closed category set) classifiers trained with supervised fine-tuning. In this paper we show, for the first time, how to reduce this gap without any labels and without any paired VL data, using an unlabeled image collection and a set of texts auto-generated using a Large Language Model (LLM) describing the categories of interest and effectively substituting labeled visual instances of those categories. Using our label-free approach, we are able to attain significant performance improvements over the zero-shot performance of the base VL model and other contemporary methods and baselines on a wide variety of datasets, demonstrating absolute improvement of up to $11.7\%$ ($3.8\%$ on average) in the label-free setting. Moreover, despite our approach being label-free, we observe $1.3\%$ average gains over leading few-shot prompting baselines that do use 5-shot supervision.

Pre-training on massive video datasets has become essential to achieve high action recognition performance on smaller downstream datasets. However, most large-scale video datasets contain images of people and hence are accompanied with issues related to privacy, ethics, and data protection, often preventing them from being publicly shared for reproducible research. Existing work has attempted to alleviate these problems by blurring faces, downsampling videos, or training on synthetic data. On the other hand, analysis on the {\em transferability} of privacy-preserving pre-trained models to downstream tasks has been limited. In this work, we study this problem by first asking the question: can we pre-train models for human action recognition with data that does not include real humans? To this end, we present, for the first time, a benchmark that leverages real-world videos with {\em humans removed} and synthetic data containing virtual humans to pre-train a model. We then evaluate the transferability of the representation learned on this data to a diverse set of downstream action recognition benchmarks. Furthermore, we propose a novel pre-training strategy, called Privacy-Preserving MAE-Align, to effectively combine synthetic data and human-removed real data. Our approach outperforms previous baselines by up to 5\% and closes the performance gap between human and no-human action recognition representations on downstream tasks, for both linear probing and fine-tuning. Our benchmark, code, and models are available at https://github.com/howardzh01/PPMA.

The ability to generalize learned representations across significantly different visual domains, such as between real photos, clipart, paintings, and sketches, is a fundamental capacity of the human visual system. In this paper, different from most cross-domain works that utilize some (or full) source domain supervision, we approach a relatively new and very practical Unsupervised Domain Generalization (UDG) setup of having no training supervision in neither source nor target domains. Our approach is based on self-supervised learning of a Bridge Across Domains (BrAD) - an auxiliary bridge domain accompanied by a set of semantics preserving visual (image-to-image) mappings to BrAD from each of the training domains. The BrAD and mappings to it are learned jointly (end-to-end) with a contrastive self-supervised representation model that semantically aligns each of the domains to its BrAD-projection, and hence implicitly drives all the domains (seen or unseen) to semantically align to each other. In this work, we show how using an edge-regularized BrAD our approach achieves significant gains across multiple benchmarks and a range of tasks, including UDG, Few-shot UDA, and unsupervised generalization across multi-domain datasets (including generalization to unseen domains and classes).

Pre-training models on Imagenet or other massive datasets of real images has led to major advances in Computer vision, albeit accompanied with shortcomings related to curation cost, privacy, usage rights, and ethical issues. In this paper, for the first time, we study the transferability of pre-trained models based on synthetic data generated by graphics simulators to downstream tasks from very different domains. In using such synthetic data for pre-training, we find that downstream performance on different tasks are fa-vored by different configurations of simulation parameters (e.g. lighting, object pose, backgrounds, etc.), and that there is no one-size-fits-all solution. It is thus better to tailor syn-thetic pre-training data to a specific downstream task, for best performance. We introduce Task2Sim, a unified model mapping downstream task representations to optimal sim-ulation parameters to generate synthetic pre-training data for them. Task2Sim learns this mapping by training to find the set of best parameters on a set of “seen” tasks. Once trained, it can then be used to predict best simulation pa-rameters for novel “unseen” tasks in one shot, without re-quiring additional training. Given a budget in number of images per class, our extensive experiments with 20 di-verse downstream tasks show Task2Sim's task-adaptive pre-training data results in significantly better downstream per-formance than non-adaptively choosing simulation param-eters on both seen and unseen tasks. It is even competitive with pre-training on real images from Imagenet.

As machine learning algorithms grow in popularity and diversify to many industries, ethical and legal concerns regarding their fairness have become increasingly relevant. We explore the problem of algorithmic fairness, taking an information-theoretic view. The maximal correlation framework is introduced for expressing fairness constraints and shown to be capable of deriving regularizers that enforce independence and separation-based fairness criteria, which admit optimization algorithms that are more computationally efficient than existing algorithms. We show that these algorithms provide smooth performance-fairness tradeoff curves and perform competitively with state-of-the-art methods on the Communities and Crimes dataset.

Foundational Models (FMs) have demonstrated unprecedented capabilities including zero-shot learning, high fidelity data synthesis, and out of domain generalization. However, the parameter capacity of FMs is still limited, leading to poor out-of-the-box performance of FMs on many expert tasks (e.g. retrieval of car manuals technical illustrations from language queries), data for which is either unseen or belonging to a long-tail part of the data distribution of the huge datasets used for FM pre-training. This underlines the necessity to explicitly evaluate and finetune FMs on such expert tasks, arguably ones that appear the most in practical real-world applications. In this paper, we propose a first of its kind FETA benchmark built around the task of teaching FMs to understand technical documentation, via learning to match their graphical illustrations to corresponding language descriptions. Our FETA benchmark focuses on text-to-image and image-to-text retrieval in public car manuals and sales catalogue brochures. FETA is equipped with a procedure for completely automatic annotation extraction (code would be released upon acceptance), allowing easy extension of FETA to more documentation types and application domains in the future. Our automatic annotation leads to an automated performance metric shown to be consistent with metrics computed on human-curated annotations (also released). We provide multiple baselines and analysis of popular FMs on FETA leading to several interesting findings that we believe would be very valuable to the FM community, paving the way towards real-world application of FMs for many practical expert tasks currently being `overlooked' by standard benchmarks focusing on common objects.

Recent action recognition models have achieved impressive results by integrating objects, their locations and interactions. However, obtaining dense structured annotations for each frame is tedious and time-consuming, making these methods expensive to train and less scalable. At the same time, if a small set of annotated images is available, either within or outside the domain of interest, how could we leverage these for a video downstream task? We propose a learning framework StructureViT (SViT for short), which demonstrates how utilizing the structure of a small number of images only available during training can improve a video model. SViT relies on two key insights. First, as both images and videos contain structured information, we enrich a transformer model with a set of object tokens that can be used across images and videos. Second, the scene representations of individual frames in video should ``align'' with those of still images. This is achieved via a Frame-Clip Consistency loss, which ensures the flow of structured information between images and videos. We explore a particular instantiation of scene structure, namely a Hand-Object Graph, consisting of hands and objects with their locations as nodes, and physical relations of contact/no-contact as edges. SViT shows strong performance improvements on multiple video understanding tasks and datasets, including the first place in the Ego4D CVPR'22 Point of No Return Temporal Localization Challenge. For code and pretrained models, visit the project page at https://eladb3.github.io/SViT/.

Action recognition has improved dramatically with massive-scale video datasets. Yet, these datasets are accompanied with issues related to curation cost, privacy, ethics, bias, and copyright. Compared to that, only minor efforts have been devoted toward exploring the potential of synthetic video data. In this work, as a stepping stone towards addressing these shortcomings, we study the transferability of video representations learned solely from synthetically-generated video clips, instead of real data. We propose SynAPT, a novel benchmark for action recognition based on a combination of existing synthetic datasets, in which a model is pre-trained on synthetic videos rendered by various graphics simulators, and then transferred to a set of downstream action recognition datasets, containing different categories than the synthetic data. We provide an extensive baseline analysis on SynAPT revealing that the simulation-to-real gap is minor for datasets with low object and scene bias, where models pre-trained with synthetic data even outperform their real data counterparts. We posit that the gap between real and synthetic action representations can be attributed to contextual bias and static objects related to the action, instead of the temporal dynamics of the action itself. The SynAPT benchmark is available at https://github.com/mintjohnkim/SynAPT.

Few-shot detection and classification have advanced significantly in recent years. Yet, detection approaches require strong annotation (bounding boxes) both for pre-training and for adaptation to novel classes, and classification approaches rarely provide localization of objects in the scene. In this paper, we introduce StarNet - a few-shot model featuring an end-to-end differentiable non-parametric star-model detection and classification head. Through this head, the backbone is meta-trained using only image-level labels to produce good features for jointly localizing and classifying previously unseen categories of few-shot test tasks using a star-model that geometrically matches between the query and support images (to find corresponding object instances). Being a few-shot detector, StarNet does not require any bounding box annotations, neither during pre-training nor for novel classes adaptation. It can thus be applied to the previously unexplored and challenging task of Weakly Supervised Few-Shot Object Detection (WS-FSOD), where it attains significant improvements over the baselines. In addition, StarNet shows significant gains on few-shot classification benchmarks that are less cropped around the objects (where object localization is key).

Few-shot learning methods offer pre-training techniques optimized for easier later adaptation of the model to new classes (unseen during training) using one or a few examples. This adaptivity to unseen classes is especially important for many practical applications where the pre-trained label space cannot remain fixed for effective use and the model needs to be "specialized" to support new categories on the fly. One particularly interesting scenario, essentially overlooked by the few-shot literature, is Coarse-to-Fine Few-Shot (C2FS), where the training classes (e.g. animals) are of much ‘coarser granularity’ than the target (test) classes (e.g. breeds). A very practical example of C2FS is when the target classes are sub-classes of the training classes. Intuitively, it is especially challenging as (both regular and few-shot) supervised pre-training tends to learn to ignore intra-class variability which is essential for separating sub-classes. In this paper, we introduce a novel ’Angular normalization’ module that allows to effectively combine supervised and self-supervised contrastive pre-training to approach the proposed C2FS task, demonstrating significant gains in a broad study over multiple baselines and datasets. We hope that this work will help to pave the way for future research on this new, challenging, and very practical topic of C2FS classification.

Nowadays, there is an abundance of data involving images and surrounding free-form text weakly corresponding to those images. Weakly Supervised phrase-Grounding (WSG) deals with the task of using this data to learn to localize (or to ground) arbitrary text phrases in images without any additional annotations. However, most recent SotA methods for WSG assume an existence of a pre-trained object detector, relying on it to produce the ROIs for localization. In this work, we focus on the task of Detector-Free WSG (DF-WSG) to solve WSG without relying on a pre-trained detector. The key idea behind our proposed Grounding by Separation (GbS) method is synthesizing ‘text to image-regions’ associations by random alpha-blending of arbitrary image pairs and using the corresponding texts of the pair as conditions to recover the alpha map from the blended image via a segmentation network. At test time, this allows using the query phrase as a condition for a non-blended query image, thus interpreting the test image as a composition of a region corresponding to the phrase and the complement region. Our GbS shows an 8.5% accuracy improvement over previous DF-WSG SotA, for a range of benchmarks including Flickr30K, Visual Genome, and ReferIt, as well as a complementary improvement (above 7%) over the detector-based approaches for WSG.

Tremendous progress has been made in visual representation learning, notably with the recent success of self-supervised contrastive learning methods. Supervised contrastive learning has also been shown to outperform its cross-entropy counterparts by leveraging labels for choosing where to contrast. However, there has been little work to explore the transfer capability of contrastive learning to a different domain. In this paper, we conduct a comprehensive study on the transferability of learned representations of different contrastive approaches for linear evaluation, full-network transfer, and few-shot recognition on 12 downstream datasets from different domains, and object detection tasks on MSCOCO and VOC0712. The results show that the contrastive approaches learn representations that are easily transferable to a different downstream task. We further observe that the joint objective of self-supervised contrastive loss with cross-entropy/supervised-contrastive loss leads to better transferability of these models over their supervised counterparts. Our analysis reveals that the representations learned from the contrastive approaches contain more low/mid-level semantics than cross-entropy models, which enables them to quickly adapt to a new task. Our codes and models will be publicly available to facilitate future research on transferability of visual representations.1

Most existing works in few-shot learning rely on meta-learning the network on a large base dataset which is typically from the same domain as the target dataset. We tackle the problem of cross-domain few-shot learning where there is a large shift between the base and target domain. The problem of cross-domain few-shot recognition with unlabeled target data is largely unaddressed in the literature. STARTUP was the first method that tackles this problem using self-training. However, it uses a fixed teacher pretrained on a labeled base dataset to create soft labels for the unlabeled target samples. As the base dataset and unlabeled dataset are from different domains, projecting the target images in the class-domain of the base dataset with a fixed pretrained model might be sub-optimal. We propose a simple dynamic distillation-based approach to facilitate unlabeled images from the novel/base dataset. We impose consistency regularization by calculating predictions from the weakly-augmented versions of the unlabeled images from a teacher network and matching it with the strongly augmented versions of the same images from a student network. The parameters of the teacher network are updated as exponential moving average of the parameters of the student network. We show that the proposed network learns representation that can be easily adapted to the target domain even though it has not been trained with target-specific classes during the pretraining phase. Our model outperforms the current state-of-the art method by 4.4% for 1-shot and 3.6% for 5-shot classification in the BSCD-FSL benchmark, and also shows competitive performance on traditional in-domain few-shot learning task.

Example synthesis is one of the leading methods to tackle the problem of few-shot learning, where only a small number of samples per class are available. However, current synthesis approaches only address the scenario of a single category label per image. In this work, we propose a novel technique for synthesizing samples with multiple labels for the (yet unhandled) multi-label few-shot classification scenario. We propose to combine pairs of given examples in feature space, so that the resulting synthesized feature vectors will correspond to examples whose label sets are obtained through certain set operations on the label sets of the corresponding input pairs. Thus, our method is capable of producing a sample containing the intersection, union or set-difference of labels present in two input samples. As we show, these set operations generalize to labels unseen during training. This enables performing augmentation on examples of novel categories, thus, facilitating multi-label few-shot classifier learning. We conduct numerous experiments showing promising results for the label-set manipulation capabilities of the proposed approach, both directly (using the classification and retrieval metrics), and in the context of performing data augmentation for multi-label few-shot learning. We propose a benchmark for this new and challenging task and show that our method compares favorably to all the common baselines.

Distance metric learning (DML) has been successfully applied to object classification, both in the standard regime of rich training data and in the few-shot scenario, where each category is represented by only a few examples. In this work, we propose a new method for DML that simultaneously learns the backbone network parameters, the embedding space, and the multi-modal distribution of each of the training categories in that space, in a single end-to-end training process. Our approach outperforms state-of-the-art methods for DML-based object classification on a variety of standard fine-grained datasets. Furthermore, we demonstrate the effectiveness of our approach on the problem of few-shot object detection, by incorporating the proposed DML architecture as a classification head into a standard object detection model. We achieve the best results on the ImageNet-LOC dataset compared to strong baselines, when only a few training examples are available. We also offer the community a new episodic benchmark based on the ImageNet dataset for the few-shot object detection task.

Deep neural networks, trained with large amount of labeled data, can fail to generalize well when tested with examples from a target domain whose distribution differs from the training data distribution, referred as the source domain. It can be expensive or even infeasible to obtain required amount of labeled data in all possible domains. Unsupervised domain adaptation sets out to address this problem, aiming to learn a good predictive model for the target domain using labeled examples from the source domain but only unlabeled examples from the target domain. Domain alignment approaches this problem by matching the source and target feature distributions, and has been used as a key component in many state-of-the-art domain adaptation methods. However, matching the marginal feature distributions does not guarantee that the corresponding class conditional distributions will be aligned across the two domains. We propose co-regularized domain alignment for unsupervised domain adaptation, which constructs multiple diverse feature spaces and aligns source and target distributions in each of them individually, while encouraging that alignments agree with each other with regard to the class predictions on the unlabeled target examples. The proposed method is generic and can be used to improve any domain adaptation method which uses domain alignment. We instantiate it in the context of a recent state-of-the-art method and observe that it provides significant performance improvements on several domain adaptation benchmarks.

Learning to classify new categories based on just one or a few examples is a long-standing challenge in modern computer vision. In this work, we propose a simple yet effective method for few-shot (and one-shot) object recognition. Our approach is based on a modified auto-encoder, denoted delta-encoder, that learns to synthesize new samples for an unseen category just by seeing few examples from it. The synthesized samples are then used to train a classifier. The proposed approach learns to both extract transferable intra-class deformations, or "deltas", between same-class pairs of training examples, and to apply those deltas to the few provided examples of a novel class (unseen during training) in order to efficiently synthesize samples from that new class. The proposed method improves the state-of-the-art of one-shot object-recognition and performs comparably in the few-shot case.

We approach the problem of fast detection and recognition of a large number (thousands) of object categories while training on a very limited amount of examples, usually one per category. Examples of this task include: (i) detection of retail products, where we have only one studio image of each product available for training, (ii) detection of brand logos, and (iii) detection of 3D objects and their respective poses within a static 2D image, where only a sparse subset of (partial) object views is available for training, with a single example for each view. Building a detector based on so few examples presents a significant challenge for the current top-performing (deep) learning based techniques, which require large amounts of data to train. Our approach for this task is based on a non-parametric probabilistic model for initial detection, CNN-based refinement and temporal integration where applicable. We successfully demonstrate its usefulness in a variety of experiments on both existing and our own benchmarks achieving state-of-the-art performance.

Detecting an object part relies on two sources of information - the appearance of the part itself and the context supplied by surrounding parts. In this paper we consider problems in which a target part cannot be recognized reliably using its own appearance, such as detecting low-resolution hands, and must be recognized using the context of surrounding parts. We develop the `chains model' which can locate parts of interest in a robust and precise manner, even when the surrounding context is highly variable and deformable. In the proposed model, the relation between context features and the target part is modeled in a non-parametric manner using an ensemble of feature chains leading from parts in the context to the detection target. The method uses the configuration of the features in the image directly rather than through fitting an articulated 3-D model of the object. In addition, the chains are composable, meaning that new chains observed in the test image can be composed of sub-chains seen during training. Consequently, the model is capable of handling object poses which are infrequent, even non-existent, during training. We test the approach in different settings, including object parts detection, as well as complete object detection. The results show the advantages of the chains model for detecting and localizing parts of complex deformable objects.

This paper presents an approach to the visual recognition of human actions using only single images as input. The task is easy for humans but difficult for current approaches to object recognition, because action instances may be similar in terms of body pose, and often require detailed examination of relations between participating objects and body parts in order to be recognized. The proposed approach applies a two-stage interpretation procedure to each training and test image. The first stage produces accurate detection of the relevant body parts of the actor, forming a prior for the local evidence needed to be considered for identifying the action. The second stage extracts features that are ‘anchored’ to the detected body parts, and uses these features and their feature-to-part relations in order to recognize the action. The body anchored priors we propose apply to a large range of human actions. These priors allow focusing on the relevant regions and relations, thereby significantly simplifying the learning process and increasing recognition performance.

Class learning, both supervised and unsupervised, requires feature selection, which includes two main components. The first is the selection of a discriminative subset of features from a larger pool. The second is the selection of detection parameters for each feature to optimize classification performance. In this paper we present a method for the discovery of multiple classification features, their detection parameters and their consistent configurations, in the fully unsupervised setting. This is achieved by a global optimization of joint consistency between the features as a function of the detection parameters, without assuming any prior parametric model. We demonstrate how the proposed framework can be applied for learning different types of feature parameters, such as detection thresholds and geometric relations, resulting in the unsupervised discovery of informative configurations of objects parts. We test our approach on a wide range of classes and show good results. We also demonstrate how the approach can be used to unsupervisedly separate and learn visually similar sub-classes of a single category, such as facial views or hand poses. We use the approach to compare various criteria for feature consistency, including Mutual Information, Suspicious Coincidence, L2 and Jaccard index. Finally, we compare our approach to aparametric consistency optimization technique such as pLSA and show significantly better performance.