Abstract The assessment of evaluation metrics (meta-evaluation) is crucial for determining the suitability of existing metrics in text-to-image (T2I) generation tasks. Human-based meta-evaluation is costly and time-intensive, and automated alternatives are scarce. We address this gap and propose CROC: a scalable framework for automated Contrastive Robustness Checks that systematically probes and quantifies metric robustness by synthesizing contrastive test cases across a comprehensive taxonomy of image properties. With CROC, we generate a pseudo-labeled dataset (CROCsyn) of over 1 million contrastive prompt–image pairs to enable a fine-grained comparison of evaluation metrics. We also use this dataset to train CROCScore, a new metric that achieves state-of-the-art performance among open-source methods, demonstrating an additional key application of our framework. To complement this dataset, we introduce a human-supervised benchmark (CROChum) targeting especially challenging categories. Our results highlight robustness issues in existing metrics: for example, many fail on prompts involving negation, and all tested open-source metrics fail on at least 24% of cases involving correct identification of body parts.1

Decoder-only large language models typically rely solely on masked causal attention, which limits their expressiveness by restricting information flow to one direction.We propose Bitune, a method that enhances pretrained decoder-only LLMs by incorporating bidirectional attention into prompt processing.We evaluate Bitune in instruction-tuning and question-answering settings, showing significant improvements in performance on commonsense reasoning, arithmetic, and language understanding tasks.Furthermore, extensive ablation studies validate the role of each component of the method, and demonstrate that Bitune is compatible with various parameterefficient finetuning techniques and full model finetuning.

Dense self-supervised learning has shown great promise for learning pixel- and patch-level representations, but extending it to videos remains challenging due to the complexity of motion dynamics. Existing approaches struggle as they rely on static augmentations that fail under object deformations, occlusions, and camera movement, leading to inconsistent feature learning over time. We propose a motion-guided self-supervised learning framework that clusters dense point tracks to learn spatiotemporally consistent representations. By leveraging an off-the-shelf point tracker, we extract long-range motion trajectories and optimize feature clustering through a momentum-encoder-based optimal transport mechanism. To ensure temporal coherence, we propagate cluster assignments along tracked points, enforcing feature consistency across views despite viewpoint changes. Integrating motion as an implicit supervisory signal, our method learns representations that generalize across frames, improving robustness in dynamic scenes and challenging occlusion scenarios. By initializing from strong image-pretrained models and leveraging video data for training, we improve state-of-the-art by 1% to 6% on six image and video datasets and four evaluation benchmarks. The implementation is publicly available at our GitHub repository: https://github.com/SMSD75/MoSiC/tree/main

Open-Vocabulary Segmentation (OVS) methods are capable of performing semantic segmentation without relying on a fixed vocabulary, and in some cases, without training or fine-tuning. However, OVS methods typically require a human in the loop to specify the vocabulary based on the task or dataset at hand. In this paper, we introduce Auto-Vocabulary Semantic Segmentation (AVS), advancing open-ended image understanding by eliminating the necessity to predefine object categories for segmentation. Our approach, AutoSeg, presents a framework that autonomously identifies relevant class names using semantically enhanced BLIP embeddings and segments them afterwards. Given that open-ended object category predictions cannot be directly compared with a fixed ground truth, we develop a Large Language Model-based Auto-Vocabulary Evaluator (LAVE) to efficiently evaluate the automatically generated classes and their corresponding segments. With AVS, our method sets new benchmarks on datasets PASCAL VOC, Context, ADE20K, and Cityscapes, while showing competitive performance to OVS methods that require specified class names.

We address the challenge of representing long captions in vision-language models, such as CLIP. By design these models are limited by fixed, absolute positional encodings, restricting inputs to a maximum of 77 tokens and hindering performance on tasks requiring longer descriptions. Although recent work has attempted to overcome this limit, their proposed approaches struggle to model token relationships over longer distances and simply extend to a fixed new token length. Instead, we propose a generalizable method, named TULIP, able to upgrade the token length to any length for CLIP-like models. We do so by improving the architecture with relative position encodings, followed by a training procedure that (i) distills the original CLIP text encoder into an encoder with relative position encodings and (ii) enhances the model for aligning longer captions with images. By effectively encoding captions longer than the default 77 tokens, our model outperforms baselines on cross-modal tasks such as retrieval and text-to-image generation. The code repository is available at https://github.com/ivonajdenkoska/tulip.

We introduce NeCo: Patch Neighbor Consistency, a novel self-supervised training loss that enforces patch-level nearest neighbor consistency across a student and teacher model. Compared to contrastive approaches that only yield binary learning signals, i.e. "attract" and "repel", this approach benefits from the more fine-grained learning signal of sorting spatially dense features relative to reference patches. Our method leverages differentiable sorting applied on top of pretrained representations, such as DINOv2-registers to bootstrap the learning signal and further improve upon them. This dense post-pretraining leads to superior performance across various models and datasets, despite requiring only 19 hours on a single GPU. This method generates high-quality dense feature encoders and establishes several new state-of-the-art results such as +2.3 % and +4.2% for non-parametric in-context semantic segmentation on ADE20k and Pascal VOC, +1.6% and +4.8% for linear segmentation evaluations on COCO-Things and -Stuff and improvements in the 3D understanding of multi-view consistency on SPair-71k, by more than 1.5%.

Self-supervised learning (SSL) has achieved major advances in natural images and video understanding, but challenges remain in domains like echocardiography (heart ultrasound) due to subtle anatomical structures, complex temporal dynamics, and the current lack of domain-specific pre-trained models. Existing SSL approaches such as contrastive, masked modeling, and clustering-based methods struggle with high intersample similarity, sensitivity to low PSNR inputs common in ultrasound, or aggressive augmentations that distort clinically relevant features. We present DISCOVR (Distilled Image Supervision for Cross Modal Video Representation), a self-supervised dual-branch framework for cardiac ultrasound video representation learning. DISCOVR combines a clustering-based video encoder that models temporal dynamics with an online image encoder that extracts fine-grained spatial semantics. These branches are connected through a semantic cluster distillation loss that transfers anatomical knowledge from the evolving image encoder to the video encoder, enabling temporally coherent representations enriched with fine-grained semantic understanding. Evaluated on six echocardiography datasets spanning fetal, pediatric, and adult populations, DISCOVR outperforms both specialized video anomaly detection methods and state-of-the-art video-SSL baselines in zero-shot and linear probing setups, achieving superior segmentation transfer and strong downstream performance on clinically relevant tasks such as LVEF prediction. **Code available at:** [https://github.com/mdivyanshu97/DISCOVR](https://github.com/mdivyanshu97/DISCOVR)

Vision foundation models achieve remarkable performance but are only available in a limited set of pre-determined sizes, forcing sub-optimal deployment choices under real-world constraints. We introduce SnapViT: single-shot network approximation for pruned Vision Transformers, a new post-pretraining structured pruning method that enables elastic inference across a continuum of compute budgets. Our approach efficiently combines gradient information with cross-network structure correlations, approximated via an evolutionary algorithm, does not require labeled data, generalizes to models without a classification head, and is retraining-free. Experiments on DINO, SigLIPv2, DeIT, and AugReg models demonstrate superior performance over state-of-the-art methods across various sparsities, requiring less than five minutes on a single A100 GPU to generate elastic models that can be adjusted to any computational budget. Our key contributions include an efficient pruning strategy for pretrained Vision Transformers, a novel evolutionary approximation of Hessian off-diagonal structures, and a self-supervised importance scoring mechanism that maintains strong performance without requiring retraining or labels. Code and pruned models are available at: https://elastic.ashita.nl/

Novel View Synthesis (NVS), which tries to produce a realistic image at the target view given source view images and their corresponding poses, is a fundamental problem in 3D Vision. As this task is heavily under-constrained, some recent work, like Zerol23 [18], tries to solve this problem with generative modeling, specifically using pre-trained diffusion models. Although this strategy generalizes well to new scenes, compared to neural radiance field-based methods, it offers low levels of flexibility. For example, it can only accept a single-view image as input, despite realistic applications often offering multiple input images. This is because the source-view images and corresponding poses are processed separately and injected into the model at different stages. Thus it is not trivial to generalize the model into multi-view source images, once they are available. To solve this issue, we try to process each pose image pair separately and then fuse them as a unified visual representation which will be injected into the model to guide image synthesis at the target-views. However, inconsistency and computation costs increase as the number of input source-view images increases. To solve these issues, the Multi-view Cross Former module is proposed which maps variable-length input data to fix-size output data. A two-stage training strategy is introduced to further improve the efficiency during training time. Qualitative and quantitative evaluation over multiple datasets demonstrates the effectiveness of the proposed method against previous approaches. The code will be released according to the acceptance.

Self-Supervised Learning (SSL) is a valuable and robust training methodology for contemporary Deep Neural Networks (DNNs), enabling unsupervised pretraining on a ‘pretext task’ that does not require ground-truth labels/annotation. This allows efficient representation learning from massive amounts of unlabeled training data, which in turn leads to increased accuracy in a ‘downstream task’ by exploiting supervised transfer learning. Despite the relatively straightforward conceptualization and applicability of SSL, it is not always feasible to collect and/or to utilize very large pretraining datasets, especially when it comes to real-world application settings. In particular, in cases of specialized and domain-specific application scenarios, it may not be achievable or practical to assemble a relevant image pretraining dataset in the order of millions of instances or it could be computationally infeasible to pretrain at this scale, e.g., due to unavailability of sufficient computational resources that SSL methods typically require to produce improved visual analysis results. This situation motivates an investigation on the effectiveness of common SSL pretext tasks, when the pretraining dataset is of relatively limited/constrained size. This work briefly introduces the main families of modern visual SSL methods and, subsequently, conducts a thorough comparative experimental evaluation in the low-data regime, targeting to identify: (a) what is learnt via low-data SSL pretraining, and (b) how do different SSL categories behave in such training scenarios. Interestingly, for domain-specific downstream tasks, in-domain low-data SSL pretraining outperforms the common approach of large-scale pretraining on general datasets. Grounded on the obtained results, valuable insights are highlighted regarding the performance of each category of SSL methods, which in turn suggest straightforward future research directions in the field.

This work proposes SAMSelect, an algorithm to obtain a salient three-channel visualization for multispectral images. We develop SAMSelect and show its use for marine scientists visually interpreting floating marine debris in Sentinel-2 imagery. These debris are notoriously difficult to visualize due to their compositional heterogeneity in medium-resolution imagery. Out of these difficulties, a visual interpretation of imagery showing marine debris remains a common practice by domain experts, who select bands and spectral indices on a case-by-case basis informed by common practices and heuristics. SAMSelect selects the band or index combination that achieves the best classification accuracy on a small annotated dataset through the Segment Anything Model. Its central assumption is that the three-channel visualization achieves the most accurate segmentation results also provide good visual information for photo-interpretation. We evaluate SAMSelect in three Sentinel-2 scenes containing generic marine debris in Accra, Ghana, and Durban, South Africa, and deployed plastic targets from the Plastic Litter Project. This reveals the potential of new previously unused band combinations (e.g., a normalized difference index of B8, B2), which demonstrate improved performance compared to literature-based indices. We describe the algorithm in this paper and provide an open-source code repository that will be helpful for domain scientists doing visual photo interpretation, especially in the marine field.

The pandemic in 2020 and 2021 had enormous economic and societal consequences, and studies show that contact tracing algorithms can be key in the early containment of the virus. While large strides have been made towards more effective contact tracing algorithms, we argue that privacy concerns currently hold deployment back. The essence of a contact tracing algorithm constitutes the communication of a risk score. Yet, it is precisely the communication and release of this score to a user that an adversary can leverage to gauge the private health status of an individual. We pinpoint a realistic attack scenario and propose a contact tracing algorithm with differential privacy guarantees against this attack. The algorithm is tested on the two most widely used agent-based COVID19 simulators and demonstrates superior performance in a wide range of settings. Especially for realistic test scenarios and while releasing each risk score with epsilon=1 differential privacy, we achieve a two to ten-fold reduction in the infection rate of the virus. To the best of our knowledge, this presents the first contact tracing algorithm with differential privacy guarantees when revealing risk scores for COVID19.

Foundation Models (FMs) have revolutionized machine learning with their adaptability and high performance across tasks; yet, their integration into Federated Learning (FL) is challenging due to substantial communication overhead from their extensive parameterization. Current communication-efficient FL strategies, such as gradient compression, reduce bitrates to around 1 bit-per-parameter (bpp). However, these approaches fail to harness the characteristics of FMs, with their large number of parameters still posing a challenge to communication efficiency, even at these bitrate regimes. In this work, we present DeltaMask, a novel method that efficiently fine-tunes FMs in FL at an ultra-low bitrate, well below 1 bpp. DeltaMask employs stochastic masking to detect highly effective subnetworks within FMs and leverage stochasticity and sparsity in client masks to compress updates into a compact grayscale image using probabilistic filters, deviating from traditional weight training approaches. Our comprehensive evaluations across various datasets and architectures demonstrate DeltaMask efficiently achieves bitrates as low as 0.09 bpp, enhancing communication efficiency while maintaining FMs performance, as measured on 8 datasets and 5 pre-trained models of various network architectures.

Vision-Language Models (VLMs), such as Flamingo and GPT-4V, have shown immense potential by integrating large language models with vision systems. Nevertheless, these models face challenges in the fundamental computer vision task of object localisation, due to their training on multi-modal data containing mostly captions without explicit spa-tial grounding. While it is possible to construct custom, supervised training pipelines with bounding box annotations that integrate with VLMs, these result in specialized and hard-to-scale models. In this paper, we aim to explore the limits of caption-based VLMs and instead propose to tackle the challenge in a simpler manner by i) keeping the weights of a caption-based VLM frozen and ii) not using any supervised detection data. To this end, we introduce an input-agnostic Positional Insert (PIN), a learnable spa-tial prompt, containing a minimal set of parameters that are slid inside the frozen VLM, unlocking object localisation capabilities. Our PIN module is trained with a simple next-token prediction task on synthetic data without requiring the introduction of new output heads. Our experiments demonstrate strong zero-shot localisation performances on a variety of images, including Pascal VOC, COCO, LVIS, and diverse images like paintings or cartoons.

While recent supervised methods for reference-based object counting continue to improve the performance on benchmark datasets, they have to rely on small datasets due to the cost associated with manually annotating dozens of objects in images. We propose UnCounTR, a model that can learn this task without requiring any manual annotations. To this end, we construct “Self-Collages”, images with various pasted objects as training samples, that provide a rich learning signal covering arbitrary object types and counts. Our method builds on existing unsupervised representations and segmentation techniques to successfully demonstrate for the first time the ability of reference-based counting without manual supervision. Our experiments show that our method not only outperforms simple base-lines and generic models such as FasterRCNN and DETR, but also matches the performance of supervised counting models in some domains.11Code: https://github.com/lukasknobel/SelfCollages

Low-rank adapation (LoRA) is a popular method that reduces the number of trainable parameters when finetuning large language models, but still faces acute storage challenges when scaling to even larger models or deploying numerous per-user or per-task adapted models. In this work, we present Vector-based Random Matrix Adaptation (VeRA), which significantly reduces the number of trainable parameters compared to LoRA, yet maintains the same performance. It achieves this by using a single pair of low-rank matrices shared across all layers and learning small scaling vectors instead. We demonstrate its effectiveness on the GLUE and E2E benchmarks, image classification tasks, and show its application in instruction-tuning of 7B and 13B language models. Website: https://dkopi.github.io/vera

This work aims to improve the efficiency of vision transformers (ViTs). While ViTs use computationally expensive self-attention operations in every layer, we identify that these operations are highly correlated across layers -- a key redundancy that causes unnecessary computations. Based on this observation, we propose SkipAT a method to reuse self-attention computation from preceding layers to approximate attention at one or more subsequent layers. To ensure that reusing self-attention blocks across layers does not degrade the performance, we introduce a simple parametric function, which outperforms the baseline transformer's performance while running computationally faster. We show that SkipAT is agnostic to transformer architecture and is effective in image classification, semantic segmentation on ADE20K, image denoising on SIDD, and video denoising on DAVIS. We achieve improved throughput at the same-or-higher accuracy levels in all these tasks.

Self-supervised learning has unlocked the potential of scaling up pretraining to billions of images, since annotation is unnecessary. But are we making the best use of data? How more economical can we be? In this work, we attempt to answer this question by making two contributions. First, we investigate first-person videos and introduce a ``Walking Tours'' dataset. These videos are high-resolution, hours-long, captured in a single uninterrupted take, depicting a large number of objects and actions with natural scene transitions. They are unlabeled and uncurated, thus realistic for self-supervision and comparable with human learning. Second, we introduce a novel self-supervised image pretraining method tailored for learning from continuous videos. Existing methods typically adapt image-based pretraining approaches to incorporate more frames. Instead, we advocate a ``tracking to learn to recognize'' approach. Our method called DoRA, leads to attention maps that **D**isc**O**ver and t**RA**ck objects over time in an end-to-end manner, using transformer cross-attention. We derive multiple views from the tracks and use them in a classical self-supervised distillation loss. Using our novel approach, a single Walking Tours video remarkably becomes a strong competitor to ImageNet for several image and video downstream tasks.

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

Pandemics have a major impact on society and the economy. In the case of a new virus, such as COVID-19, high-grade tests and vaccines might be slow to develop and scarce in the crucial initial phase. With no time to waste and lock-downs being expensive, contact tracing is thus an essential tool for policymakers. In theory, statistical inference on a virus transmission model can provide an effective method for tracing infections. However, in practice, such algorithms need to run decentralized, rendering existing methods – that require hundreds or even thousands of daily messages per person – infeasible. In this paper, we develop an algorithm that (i) requires only a few (2-5) daily messages, (ii) works with extremely low bandwidths (3-5 bits) and (iii) enables quarantining and targeted testing that drastically reduces the peak and length of the pandemic. We compare the effectiveness of our algorithm using two agent-based simulators of realistic contact patterns and pandemic parameters and show that it performs well even with low bandwidth, imprecise tests, and incomplete population coverage.

Diffusion models have demonstrated remarkable progress in image generation quality, especially when guidance is used to control the generative process. However, guidance requires a large amount of image-annotation pairs for training and is thus dependent on their availability and correctness. In this paper, we eliminate the need for such annotation by instead exploiting the flexibility of self-supervision signals to design a framework for self-guided diffusion models. By leveraging a feature extraction function and a self-annotation function, our method provides guidance signals at various image granularities: from the level of holistic images to object boxes and even segmentation masks. Our experiments on single-label and multi-label image datasets demonstrate that self-labeled guidance always outperforms diffusion models without guidance and may even surpass guidance based on ground-truth labels. When equipped with self-supervised box or mask proposals, our method further generates visually diverse yet semantically consistent images, without the need for any class, box, or segment label annotation. Self-guided diffusion is simple, flexible and expected to profit from deployment at scale.

Spatially dense self-supervised learning is a rapidly growing problem domain with promising applications for unsupervised segmentation and pretraining for dense downstream tasks. Despite the abundance of temporal data in the form of videos, this information-rich source has been largely overlooked. Our paper aims to address this gap by proposing a novel approach that incorporates temporal consistency in dense self-supervised learning. While methods designed solely for images face difficulties in achieving even the same performance on videos, our method improves not only the representation quality for videos – but also images. Our approach, which we call time-tuning, starts from image-pretrained models and fine-tunes them with a novel self-supervised temporal-alignment clustering loss on unlabeled videos. This effectively facilitates the transfer of high-level information from videos to image representations. Time-tuning improves the state-of-the-art by 8-10% for unsupervised semantic segmentation on videos and matches it for images. We believe this method paves the way for further self-supervised scaling by leveraging the abundant availability of videos. The implementation can be found here : https://github.com/SMSD75/Timetuning

In this paper we address the task of finding representative subsets of points in a 3D point cloud by means of a point-wise ordering. Only a few works have tried to address this challenging vision problem, all with the help of hard to obtain point and cloud labels. Different from these works, we introduce the task of point-wise ordering in 3D point clouds through self-supervision, which we call selfordering. We further contribute the first end-to-end trainable network that learns a point-wise ordering in a selfsupervised fashion. It utilizes a novel differentiable point scoring-sorting strategy and it constructs an hierarchical contrastive scheme to obtain self-supervision signals. We extensively ablate the method and show its scalability and superior performance even compared to supervised ordering methods on multiple datasets and tasks including zeroshot ordering of point clouds from unseen categories.

Identifying the causal variables of an environment and how to intervene on them is of core value in applications such as robotics and embodied AI. While an agent can commonly interact with the environment and may implicitly perturb the behavior of some of these causal variables, often the targets it affects remain unknown. In this paper, we show that causal variables can still be identified for many common setups, e.g., additive Gaussian noise models, if the agent’s interactions with a causal variable can be described by an unknown binary variable. This happens when each causal variable has two different mechanisms, e.g., an observational and an interventional one. Using this identifiability result, we propose BISCUIT, a method for simultaneously learning causal variables and their corresponding binary interaction variables. On three robotic-inspired datasets, BISCUIT accurately identifies causal variables and can even be scaled to complex, realistic environments for embodied AI.

We tackle the problem of learning object detectors without supervision. Differently from weakly-supervised object detection, we do not assume image-level class labels. Instead, we extract a supervisory signal from audio-visual data, using the audio component to “teach” the object detector. While this problem is related to sound source localisation, it is considerably harder because the detector must classify the objects by type, enumerate each instance of the object, and do so even when the object is silent. We tackle this problem by first designing a self-supervised framework with a contrastive objective that jointly learns to classify and localise objects. Then, without using any supervision, we simply use these self-supervised labels and boxes to train an image-based object detector. With this, we outperform previous unsupervised and weakly-supervised detectors for the task of object detection and sound source localization. We also show that we can align this detector to ground-truth classes with as little as one label per pseudo-class, and show how our method can learn to detect generic objects that go beyond instruments, such as airplanes and cats.

Progress in self-supervised learning has brought strong image representation learning methods. Yet so far, it has mostly focused on image-level learning. In turn, tasks such as unsupervised image segmentation have not benefited from this trend as they require spatially-diverse representations. However, learning dense representations is challenging, as in the unsupervised context it is not clear how to guide the model to learn representations that correspond to various potential object categories. In this paper, we argue that self-supervised learning of object parts is a solution to this issue. Object parts are generalizable: they are a priori independent of an object definition, but can be grouped to form objects a posteriori. To this end, we leverage the recently proposed Vision Transformer's capability of attending to objects and combine it with a spatially dense clustering task for fine-tuning the spatial tokens. Our method surpasses the state-of-the-art on three semantic segmentation benchmarks by 17%-3%, showing that our representations are versatile under various object definitions. Finally, we extend this to fully unsupervised segmentation - which refrains completely from using label information even at test-time - and demonstrate that a simple method for automatically merging discovered object parts based on community detection yields substantial gains..

Understanding the latent causal factors of a dynamical system from visual observations is considered a crucial step towards agents reasoning in complex environments. In this paper, we propose CITRIS, a variational autoencoder framework that learns causal representations from temporal sequences of images in which underlying causal factors have possibly been intervened upon. In contrast to the recent literature, CITRIS exploits temporality and observing intervention targets to identify scalar and multidimensional causal factors, such as 3D rotation angles. Furthermore, by introducing a normalizing flow, CITRIS can be easily extended to leverage and disentangle representations obtained by already pretrained autoencoders. Extending previous results on scalar causal factors, we prove identifiability in a more general setting, in which only some components of a causal factor are affected by interventions. In experiments on 3D rendered image sequences, CITRIS outperforms previous methods on recovering the underlying causal variables. Moreover, using pretrained autoencoders, CITRIS can even generalize to unseen instantiations of causal factors, opening future research areas in sim-to-real generalization for causal representation learning.

The quality of the image representations obtained from self-supervised learning depends strongly on the type of data augmentations used in the learning formulation. Recent papers have ported these methods from still images to videos and found that leveraging both audio and video signals yields strong gains; however, they did not find that spatial augmentations such as cropping, which are very important for still images, work as well for videos. In this paper, we improve these formulations in two ways unique to the spatio-temporal aspect of videos. First, for space, we show that spatial augmentations such as cropping do work well for videos too, but that previous implementations, due to the high processing and memory cost, could not do this at a scale sufficient for it to work well. To address this issue, we first introduce Feature Crop, a method to simulate such augmentations much more efficiently directly in feature space. Second, we show that as opposed to naïve average pooling, the use of transformer-based attention improves performance significantly, and is well suited for processing feature crops. Combining both of our discoveries into a new method, Space-Time Crop & Attend (STiCA) we achieve state-of-the-art performance across multiple video-representation learning benchmarks. In particular, we achieve new state-of-the-art accuracies of 67.0% on HMDB-51 and 93.1% on UCF-101 when pre-training on Kinetics-400. Code and pretrained models are available1.

In the image domain, excellent representations can be learned by inducing invariance to content-preserving transformations via noise contrastive learning. In this paper, we generalize contrastive learning to a wider set of transformations, and their compositions, for which either invariance or distinctiveness is sought. We show that it is not immediately obvious how existing methods such as SimCLR can be extended to do so. Instead, we introduce a number of formal requirements that all contrastive formulations must satisfy, and propose a practical construction which satisfies these requirements. In order to maximise the reach of this analysis, we express all components of noise contrastive formulations as the choice of certain generalized transformations of the data (GDTs), including data sampling. We then consider videos as an example of data in which a large variety of transformations are applicable, accounting for the extra modalities – for which we analyze audio and text – and the dimension of time. We find that being invariant to certain transformations and distinctive to others is critical to learning effective video representations, improving the state-of-the-art for multiple benchmarks by a large margin, and even surpassing supervised pretraining. Code and pretrained models are available1.

The capabilities of natural language models trained on large-scale data have increased immensely over the past few years. Open source libraries such as HuggingFace have made these models easily available and accessible. While prior research has identified biases in large language models, this paper considers biases contained in the most popular versions of these models when applied `out-of-the-box' for downstream tasks. We focus on generative language models as they are well-suited for extracting biases inherited from training data. Specifically, we conduct an in-depth analysis of GPT-2, which is the most downloaded text generation model on HuggingFace, with over half a million downloads per month. We assess biases related to occupational associations for different protected categories by intersecting gender with religion, sexuality, ethnicity, political affiliation, and continental name origin. Using a template-based data collection pipeline, we collect 396K sentence completions made by GPT-2 and find: (i) The machine-predicted jobs are less diverse and more stereotypical for women than for men, especially for intersections; (ii) Intersectional interactions are highly relevant for occupational associations, which we quantify by fitting 262 logistic models; (iii) For most occupations, GPT-2 reflects the skewed gender and ethnicity distribution found in US Labor Bureau data, and even pulls the societally-skewed distribution towards gender parity in cases where its predictions deviate from real labor market observations. This raises the normative question of what language models \textit{should} learn - whether they should reflect or correct for existing inequalities.

In video transformers, the time dimension is often treated in the same way as the two spatial dimensions. However, in a scene where objects or the camera may move, a physical point imaged at one location in frame $t$ may be entirely unrelated to what is found at that location in frame $t+k$. These temporal correspondences should be modeled to facilitate learning about dynamic scenes. To this end, we propose a new drop-in block for video transformers - trajectory attention - that aggregates information along implicitly determined motion paths. We additionally propose a new method to address the quadratic dependence of computation and memory on the input size, which is particularly important for high resolution or long videos. While these ideas are useful in a range of settings, we apply them to the specific task of video action recognition with a transformer model and obtain state-of-the-art results on the Kinetics, Something-Something V2, and Epic-Kitchens datasets.

A large part of the current success of deep learning lies in the effectiveness of data -- more precisely: of labeled data. Yet, labelling a dataset with human annotation continues to carry high costs, especially for videos. While in the image domain, recent methods have allowed to generate meaningful (pseudo-) labels for unlabelled datasets without supervision, this development is missing for the video domain where learning feature representations is the current focus. In this work, we a) show that unsupervised labelling of a video dataset does not come for free from strong feature encoders and b) propose a novel clustering method that allows pseudo-labelling of a video dataset without any human annotations, by leveraging the natural correspondence between audio and visual modalities. An extensive analysis shows that the resulting clusters have high semantic overlap to ground truth human labels. We further introduce the first benchmarking results on unsupervised labelling of common video datasets.