Tinne Tuytelaars

Andrea Cossu, Tinne Tuytelaars, Antonio Carta et al.

Pre-trained models are nowadays a fundamental component of machine learning research. In continual learning, they are commonly used to initialize the model before training on the stream of non-stationary data. However, pre-training is rarely applied during continual learning. We formalize and investigate the characteristics of the continual pre-training scenario in both language and vision environments, where a model is continually pre-trained on a stream of incoming data and only later fine-tuned to different downstream tasks. We show that continually pre-trained models are robust against catastrophic forgetting and we provide strong empirical evidence supporting the fact that self-supervised pre-training is more effective in retaining previous knowledge than supervised protocols. Code is provided at https://github.com/AndreaCossu/continual-pretraining-nlp-vision .

Eli Verwimp, Rahaf Aljundi, Shai Ben-David et al. · deepmind

Continual learning is a subfield of machine learning, which aims to allow machine learning models to continuously learn on new data, by accumulating knowledge without forgetting what was learned in the past. In this work, we take a step back, and ask: "Why should one care about continual learning in the first place?". We set the stage by examining recent continual learning papers published at four major machine learning conferences, and show that memory-constrained settings dominate the field. Then, we discuss five open problems in machine learning, and even though they might seem unrelated to continual learning at first sight, we show that continual learning will inevitably be part of their solution. These problems are model editing, personalization and specialization, on-device learning, faster (re-)training and reinforcement learning. Finally, by comparing the desiderata from these unsolved problems and the current assumptions in continual learning, we highlight and discuss four future directions for continual learning research. We hope that this work offers an interesting perspective on the future of continual learning, while displaying its potential value and the paths we have to pursue in order to make it successful. This work is the result of the many discussions the authors had at the Dagstuhl seminar on Deep Continual Learning, in March 2023.

Zifu Wang, Maxim Berman, Amal Rannen-Triki et al.

Semantic segmentation datasets often exhibit two types of imbalance: \textit{class imbalance}, where some classes appear more frequently than others and \textit{size imbalance}, where some objects occupy more pixels than others. This causes traditional evaluation metrics to be biased towards \textit{majority classes} (e.g. overall pixel-wise accuracy) and \textit{large objects} (e.g. mean pixel-wise accuracy and per-dataset mean intersection over union). To address these shortcomings, we propose the use of fine-grained mIoUs along with corresponding worst-case metrics, thereby offering a more holistic evaluation of segmentation techniques. These fine-grained metrics offer less bias towards large objects, richer statistical information, and valuable insights into model and dataset auditing. Furthermore, we undertake an extensive benchmark study, where we train and evaluate 15 modern neural networks with the proposed metrics on 12 diverse natural and aerial segmentation datasets. Our benchmark study highlights the necessity of not basing evaluations on a single metric and confirms that fine-grained mIoUs reduce the bias towards large objects. Moreover, we identify the crucial role played by architecture designs and loss functions, which lead to best practices in optimizing fine-grained metrics. The code is available at \href{https://github.com/zifuwanggg/JDTLosses}{https://github.com/zifuwanggg/JDTLosses}.

Gorjan Radevski, Dusan Grujicic, Marie-Francine Moens et al.

The focal point of egocentric video understanding is modelling hand-object interactions. Standard models, e.g. CNNs or Vision Transformers, which receive RGB frames as input perform well. However, their performance improves further by employing additional input modalities that provide complementary cues, such as object detections, optical flow, audio, etc. The added complexity of the modality-specific modules, on the other hand, makes these models impractical for deployment. The goal of this work is to retain the performance of such a multimodal approach, while using only the RGB frames as input at inference time. We demonstrate that for egocentric action recognition on the Epic-Kitchens and the Something-Something datasets, students which are taught by multimodal teachers tend to be more accurate and better calibrated than architecturally equivalent models trained on ground truth labels in a unimodal or multimodal fashion. We further adopt a principled multimodal knowledge distillation framework, allowing us to deal with issues which occur when applying multimodal knowledge distillation in a naive manner. Lastly, we demonstrate the achieved reduction in computational complexity, and show that our approach maintains higher performance with the reduction of the number of input views. We release our code at https://github.com/gorjanradevski/multimodal-distillation.

Thomas Stegmüller, Tim Lebailly, Behzad Bozorgtabar et al.

Learning dense visual representations without labels is an arduous task and more so from scene-centric data. We propose to tackle this challenging problem by proposing a Cross-view consistency objective with an Online Clustering mechanism (CrOC) to discover and segment the semantics of the views. In the absence of hand-crafted priors, the resulting method is more generalizable and does not require a cumbersome pre-processing step. More importantly, the clustering algorithm conjointly operates on the features of both views, thereby elegantly bypassing the issue of content not represented in both views and the ambiguous matching of objects from one crop to the other. We demonstrate excellent performance on linear and unsupervised segmentation transfer tasks on various datasets and similarly for video object segmentation. Our code and pre-trained models are publicly available at https://github.com/stegmuel/CrOC.

Tingyu Qu, Tinne Tuytelaars, Marie-Francine Moens

News Image Captioning aims to create captions from news articles and images, emphasizing the connection between textual context and visual elements. Recognizing the significance of human faces in news images and the face-name co-occurrence pattern in existing datasets, we propose a face-naming module for learning better name embeddings. Apart from names, which can be directly linked to an image area (faces), news image captions mostly contain context information that can only be found in the article. We design a retrieval strategy using CLIP to retrieve sentences that are semantically close to the image, mimicking human thought process of linking articles to images. Furthermore, to tackle the problem of the imbalanced proportion of article context and image context in captions, we introduce a simple yet effective method Contrasting with Language Model backbone (CoLaM) to the training pipeline. We conduct extensive experiments to demonstrate the efficacy of our framework. We out-perform the previous state-of-the-art (without external data) by 7.97/5.80 CIDEr scores on GoodNews/NYTimes800k. Our code is available at https://github.com/tingyu215/VACNIC.

Tim Lebailly, Thomas Stegmüller, Behzad Bozorgtabar et al.

Leveraging nearest neighbor retrieval for self-supervised representation learning has proven beneficial with object-centric images. However, this approach faces limitations when applied to scene-centric datasets, where multiple objects within an image are only implicitly captured in the global representation. Such global bootstrapping can lead to undesirable entanglement of object representations. Furthermore, even object-centric datasets stand to benefit from a finer-grained bootstrapping approach. In response to these challenges, we introduce a novel Cross-Image Object-Level Bootstrapping method tailored to enhance dense visual representation learning. By employing object-level nearest neighbor bootstrapping throughout the training, CrIBo emerges as a notably strong and adequate candidate for in-context learning, leveraging nearest neighbor retrieval at test time. CrIBo shows state-of-the-art performance on the latter task while being highly competitive in more standard downstream segmentation tasks. Our code and pretrained models are publicly available at https://github.com/tileb1/CrIBo.

Mingxiao Li, Zehao Wang, Tinne Tuytelaars et al.

In this work, we study the problem of Embodied Referring Expression Grounding, where an agent needs to navigate in a previously unseen environment and localize a remote object described by a concise high-level natural language instruction. When facing such a situation, a human tends to imagine what the destination may look like and to explore the environment based on prior knowledge of the environmental layout, such as the fact that a bathroom is more likely to be found near a bedroom than a kitchen. We have designed an autonomous agent called Layout-aware Dreamer (LAD), including two novel modules, that is, the Layout Learner and the Goal Dreamer to mimic this cognitive decision process. The Layout Learner learns to infer the room category distribution of neighboring unexplored areas along the path for coarse layout estimation, which effectively introduces layout common sense of room-to-room transitions to our agent. To learn an effective exploration of the environment, the Goal Dreamer imagines the destination beforehand. Our agent achieves new state-of-the-art performance on the public leaderboard of the REVERIE dataset in challenging unseen test environments with improvement in navigation success (SR) by 4.02% and remote grounding success (RGS) by 3.43% compared to the previous state-of-the-art. The code is released at https://github.com/zehao-wang/LAD

Tim Lebailly, Tinne Tuytelaars

The downstream accuracy of self-supervised methods is tightly linked to the proxy task solved during training and the quality of the gradients extracted from it. Richer and more meaningful gradients updates are key to allow self-supervised methods to learn better and in a more efficient manner. In a typical self-distillation framework, the representation of two augmented images are enforced to be coherent at the global level. Nonetheless, incorporating local cues in the proxy task can be beneficial and improve the model accuracy on downstream tasks. This leads to a dual objective in which, on the one hand, coherence between global-representations is enforced and on the other, coherence between local-representations is enforced. Unfortunately, an exact correspondence mapping between two sets of local-representations does not exist making the task of matching local-representations from one augmentation to another non-trivial. We propose to leverage the spatial information in the input images to obtain geometric matchings and compare this geometric approach against previous methods based on similarity matchings. Our study shows that not only 1) geometric matchings perform better than similarity based matchings in low-data regimes but also 2) that similarity based matchings are highly hurtful in low-data regimes compared to the vanilla baseline without local self-distillation. The code is available at https://github.com/tileb1/global-local-self-distillation.

Tim Lebailly, Thomas Stegmüller, Behzad Bozorgtabar et al.

Most self-supervised methods for representation learning leverage a cross-view consistency objective i.e., they maximize the representation similarity of a given image's augmented views. Recent work NNCLR goes beyond the cross-view paradigm and uses positive pairs from different images obtained via nearest neighbor bootstrapping in a contrastive setting. We empirically show that as opposed to the contrastive learning setting which relies on negative samples, incorporating nearest neighbor bootstrapping in a self-distillation scheme can lead to a performance drop or even collapse. We scrutinize the reason for this unexpected behavior and provide a solution. We propose to adaptively bootstrap neighbors based on the estimated quality of the latent space. We report consistent improvements compared to the naive bootstrapping approach and the original baselines. Our approach leads to performance improvements for various self-distillation method/backbone combinations and standard downstream tasks. Our code is publicly available at https://github.com/tileb1/AdaSim.

Cédric Picron, Punarjay Chakravarty, Tinne Tuytelaars

Recently, two-stage Deformable DETR introduced the query-based two-stage head, a new type of two-stage head different from the region-based two-stage heads of classical detectors as Faster R-CNN. In query-based two-stage heads, the second stage selects one feature per detection processed by a transformer, called the query, as opposed to pooling a rectangular grid of features processed by CNNs as in region-based detectors. In this work, we improve the query-based head by improving the prior of the cross-attention operation with anchors, significantly speeding up the convergence while increasing its performance. Additionally, we empirically show that by improving the cross-attention prior, auxiliary losses and iterative bounding box mechanisms typically used by DETR-based detectors are no longer needed. By combining the best of both the classical and the DETR-based detectors, our FQDet head peaks at 45.4 AP on the 2017 COCO validation set when using a ResNet-50+TPN backbone, only after training for 12 epochs using the 1x schedule. We outperform other high-performing two-stage heads such as e.g. Cascade R-CNN, while using the same backbone and while being computationally cheaper. Additionally, when using the large ResNeXt-101-DCN+TPN backbone and multi-scale testing, our FQDet head achieves 52.9 AP on the 2017 COCO test-dev set after only 12 epochs of training. Code is released at https://github.com/CedricPicron/FQDet .

Georgios Kouros, Shubham Shrivastava, Cédric Picron et al.

Pose estimation is usually tackled as either a bin classification or a regression problem. In both cases, the idea is to directly predict the pose of an object. This is a non-trivial task due to appearance variations between similar poses and similarities between dissimilar poses. Instead, we follow the key idea that comparing two poses is easier than directly predicting one. Render-and-compare approaches have been employed to that end, however, they tend to be unstable, computationally expensive, and slow for real-time applications. We propose doing category-level pose estimation by learning an alignment metric in an embedding space using a contrastive loss with a dynamic margin and a continuous pose-label space. For efficient inference, we use a simple real-time image retrieval scheme with a pre-rendered and pre-embedded reference set of renderings. To achieve robustness to real-world conditions, we employ synthetic occlusions, bounding box perturbations, and appearance augmentations. Our approach achieves state-of-the-art performance on PASCAL3D and OccludedPASCAL3D and surpasses the competing methods on KITTI3D in a cross-dataset evaluation setting. The code is currently available at https://github.com/gkouros/contrastive-pose-retrieval.

Lucas Caccia, Rahaf Aljundi, Nader Asadi et al.

In the online continual learning paradigm, agents must learn from a changing distribution while respecting memory and compute constraints. Experience Replay (ER), where a small subset of past data is stored and replayed alongside new data, has emerged as a simple and effective learning strategy. In this work, we focus on the change in representations of observed data that arises when previously unobserved classes appear in the incoming data stream, and new classes must be distinguished from previous ones. We shed new light on this question by showing that applying ER causes the newly added classes' representations to overlap significantly with the previous classes, leading to highly disruptive parameter updates. Based on this empirical analysis, we propose a new method which mitigates this issue by shielding the learned representations from drastic adaptation to accommodate new classes. We show that using an asymmetric update rule pushes new classes to adapt to the older ones (rather than the reverse), which is more effective especially at task boundaries, where much of the forgetting typically occurs. Empirical results show significant gains over strong baselines on standard continual learning benchmarks

Georgios Kouros, Minye Wu, Shubham Shrivastava et al.

Neural Radiance Fields (NeRFs) have revolutionized the field of novel view synthesis, demonstrating remarkable performance. However, the modeling and rendering of reflective objects remain challenging problems. Recent methods have shown significant improvements over the baselines in handling reflective scenes, albeit at the expense of efficiency. In this work, we aim to strike a balance between efficiency and quality. To this end, we investigate an implicit-explicit approach based on conventional volume rendering to enhance the reconstruction quality and accelerate the training and rendering processes. We adopt an efficient density-based grid representation and reparameterize the reflected radiance in our pipeline. Our proposed reflection-aware approach achieves a competitive quality efficiency trade-off compared to competing methods. Based on our experimental results, we propose and discuss hypotheses regarding the factors influencing the results of density-based methods for reconstructing reflective objects. The source code is available at https://github.com/gkouros/ref-dvgo.

Jonathan Tonglet, Iryna Gurevych, Tinne Tuytelaars et al.

Identifying when and where a news image was taken is crucial for journalists and forensic experts to produce credible stories and debunk misinformation. While many existing methods rely on reverse image search (RIS) engines, these tools often fail to return results, thereby limiting their practical applicability. In this work, we address the challenging scenario where RIS evidence is unavailable. We introduce NewsRECON, a method that links images to relevant news articles to infer their date and location from article metadata. NewsRECON leverages a corpus of over 90,000 articles and integrates: (1) a bi-encoder for retrieving event-relevant articles; (2) two cross-encoders for reranking articles by location and event consistency. Experiments on the TARA and 5Pils-OOC show that NewsRECON outperforms prior work and can be combined with a multimodal large language model to achieve new SOTA results in the absence of RIS evidence. We make our code available.

Zehao Wang, Han Zhou, Matthew B. Blaschko et al.

Novel view synthesis of indoor scenes can be achieved by capturing a monocular video sequence of the environment. However, redundant information caused by artificial movements in the input video data reduces the efficiency of scene modeling. To address this, we formulate the problem as a combinatorial optimization task for view subset selection. In this work, we propose a novel subset selection framework that integrates a comprehensive diversity-based measurement with well-designed utility functions. We provide a theoretical analysis of these utility functions and validate their effectiveness through extensive experiments. Furthermore, we introduce IndoorTraj, a novel dataset designed for indoor novel view synthesis, featuring complex and extended trajectories that simulate intricate human behaviors. Experiments on IndoorTraj show that our framework consistently outperforms baseline strategies while using only 5-20% of the data, highlighting its remarkable efficiency and effectiveness. The code is available at: https://github.com/zehao-wang/IndoorTraj

Zehao Wang, Minye Wu, Yixin Cao et al. · pku

This study presents a novel evaluation framework for the Vision-Language Navigation (VLN) task. It aims to diagnose current models for various instruction categories at a finer-grained level. The framework is structured around the context-free grammar (CFG) of the task. The CFG serves as the basis for the problem decomposition and the core premise of the instruction categories design. We propose a semi-automatic method for CFG construction with the help of Large-Language Models (LLMs). Then, we induct and generate data spanning five principal instruction categories (i.e. direction change, landmark recognition, region recognition, vertical movement, and numerical comprehension). Our analysis of different models reveals notable performance discrepancies and recurrent issues. The stagnation of numerical comprehension, heavy selective biases over directional concepts, and other interesting findings contribute to the development of future language-guided navigation systems.

Liao Wang, Qiang Hu, Qihan He et al.

The success of the Neural Radiance Fields (NeRFs) for modeling and free-view rendering static objects has inspired numerous attempts on dynamic scenes. Current techniques that utilize neural rendering for facilitating free-view videos (FVVs) are restricted to either offline rendering or are capable of processing only brief sequences with minimal motion. In this paper, we present a novel technique, Residual Radiance Field or ReRF, as a highly compact neural representation to achieve real-time FVV rendering on long-duration dynamic scenes. ReRF explicitly models the residual information between adjacent timestamps in the spatial-temporal feature space, with a global coordinate-based tiny MLP as the feature decoder. Specifically, ReRF employs a compact motion grid along with a residual feature grid to exploit inter-frame feature similarities. We show such a strategy can handle large motions without sacrificing quality. We further present a sequential training scheme to maintain the smoothness and the sparsity of the motion/residual grids. Based on ReRF, we design a special FVV codec that achieves three orders of magnitudes compression rate and provides a companion ReRF player to support online streaming of long-duration FVVs of dynamic scenes. Extensive experiments demonstrate the effectiveness of ReRF for compactly representing dynamic radiance fields, enabling an unprecedented free-viewpoint viewing experience in speed and quality.

Eli Verwimp, Kuo Yang, Sarah Parisot et al.

In this paper we describe the design and the ideas motivating a new Continual Learning benchmark for Autonomous Driving (CLAD), that focuses on the problems of object classification and object detection. The benchmark utilises SODA10M, a recently released large-scale dataset that concerns autonomous driving related problems. First, we review and discuss existing continual learning benchmarks, how they are related, and show that most are extreme cases of continual learning. To this end, we survey the benchmarks used in continual learning papers at three highly ranked computer vision conferences. Next, we introduce CLAD-C, an online classification benchmark realised through a chronological data stream that poses both class and domain incremental challenges; and CLAD-D, a domain incremental continual object detection benchmark. We examine the inherent difficulties and challenges posed by the benchmark, through a survey of the techniques and methods used by the top-3 participants in a CLAD-challenge workshop at ICCV 2021. We conclude with possible pathways to improve the current continual learning state of the art, and which directions we deem promising for future research.

Matthias De Lange, Gido van de Ven, Tinne Tuytelaars

Time-dependent data-generating distributions have proven to be difficult for gradient-based training of neural networks, as the greedy updates result in catastrophic forgetting of previously learned knowledge. Despite the progress in the field of continual learning to overcome this forgetting, we show that a set of common state-of-the-art methods still suffers from substantial forgetting upon starting to learn new tasks, except that this forgetting is temporary and followed by a phase of performance recovery. We refer to this intriguing but potentially problematic phenomenon as the stability gap. The stability gap had likely remained under the radar due to standard practice in the field of evaluating continual learning models only after each task. Instead, we establish a framework for continual evaluation that uses per-iteration evaluation and we define a new set of metrics to quantify worst-case performance. Empirically we show that experience replay, constraint-based replay, knowledge-distillation, and parameter regularization methods are all prone to the stability gap; and that the stability gap can be observed in class-, task-, and domain-incremental learning benchmarks. Additionally, a controlled experiment shows that the stability gap increases when tasks are more dissimilar. Finally, by disentangling gradients into plasticity and stability components, we propose a conceptual explanation for the stability gap.

Pierre-François De Plaen, Nicola Marinello, Marc Proesmans et al.

The DEtection TRansformer (DETR) opened new possibilities for object detection by modeling it as a translation task: converting image features into object-level representations. Previous works typically add expensive modules to DETR to perform Multi-Object Tracking (MOT), resulting in more complicated architectures. We instead show how DETR can be turned into a MOT model by employing an instance-level contrastive loss, a revised sampling strategy and a lightweight assignment method. Our training scheme learns object appearances while preserving detection capabilities and with little overhead. Its performance surpasses the previous state-of-the-art by +2.6 mMOTA on the challenging BDD100K dataset and is comparable to existing transformer-based methods on the MOT17 dataset.

Henri De Plaen, Pierre-François De Plaen, Johan A. K. Suykens et al.

During training, supervised object detection tries to correctly match the predicted bounding boxes and associated classification scores to the ground truth. This is essential to determine which predictions are to be pushed towards which solutions, or to be discarded. Popular matching strategies include matching to the closest ground truth box (mostly used in combination with anchors), or matching via the Hungarian algorithm (mostly used in anchor-free methods). Each of these strategies comes with its own properties, underlying losses, and heuristics. We show how Unbalanced Optimal Transport unifies these different approaches and opens a whole continuum of methods in between. This allows for a finer selection of the desired properties. Experimentally, we show that training an object detection model with Unbalanced Optimal Transport is able to reach the state-of-the-art both in terms of Average Precision and Average Recall as well as to provide a faster initial convergence. The approach is well suited for GPU implementation, which proves to be an advantage for large-scale models.

Thomas Verelst, Paul K. Rubenstein, Marcin Eichner et al.

As natural images usually contain multiple objects, multi-label image classification is more applicable "in the wild" than single-label classification. However, exhaustively annotating images with every object of interest is costly and time-consuming. We aim to train multi-label classifiers from single-label annotations only. We show that adding a consistency loss, ensuring that the predictions of the network are consistent over consecutive training epochs, is a simple yet effective method to train multi-label classifiers in a weakly supervised setting. We further extend this approach spatially, by ensuring consistency of the spatial feature maps produced over consecutive training epochs, maintaining per-class running-average heatmaps for each training image. We show that this spatial consistency loss further improves the multi-label mAP of the classifiers. In addition, we show that this method overcomes shortcomings of the "crop" data-augmentation by recovering correct supervision signal even when most of the single ground truth object is cropped out of the input image by the data augmentation. We demonstrate gains of the consistency and spatial consistency losses over the binary cross-entropy baseline, and over competing methods, on MS-COCO and Pascal VOC. We also demonstrate improved multi-label classification mAP on ImageNet-1K using the ReaL multi-label validation set.

Shipeng Yan, Lanqing Hong, Hang Xu et al.

Vision-language pre-training (VLP) has attracted increasing attention recently. With a large amount of image-text pairs, VLP models trained with contrastive loss have achieved impressive performance in various tasks, especially the zero-shot generalization on downstream datasets. In practical applications, however, massive data are usually collected in a streaming fashion, requiring VLP models to continuously integrate novel knowledge from incoming data and retain learned knowledge. In this work, we focus on learning a VLP model with sequential chunks of image-text pair data. To tackle the catastrophic forgetting issue in this multi-modal continual learning setting, we first introduce pseudo text replay that generates hard negative texts conditioned on the training images in memory, which not only better preserves learned knowledge but also improves the diversity of negative samples in the contrastive loss. Moreover, we propose multi-modal knowledge distillation between images and texts to align the instance-wise prediction between old and new models. We incrementally pre-train our model on both the instance and class incremental splits of the Conceptual Caption dataset, and evaluate the model on zero-shot image classification and image-text retrieval tasks. Our method consistently outperforms the existing baselines with a large margin, which demonstrates its superiority. Notably, we realize an average performance boost of $4.60\%$ on image-classification downstream datasets for the class incremental split.

Minye Wu, Tinne Tuytelaars

Recent advancements in photo-realistic novel view synthesis have been significantly driven by Gaussian Splatting (3DGS). Nevertheless, the explicit nature of 3DGS data entails considerable storage requirements, highlighting a pressing need for more efficient data representations. To address this, we present Implicit Gaussian Splatting (IGS), an innovative hybrid model that integrates explicit point clouds with implicit feature embeddings through a multi-level tri-plane architecture. This architecture features 2D feature grids at various resolutions across different levels, facilitating continuous spatial domain representation and enhancing spatial correlations among Gaussian primitives. Building upon this foundation, we introduce a level-based progressive training scheme, which incorporates explicit spatial regularization. This method capitalizes on spatial correlations to enhance both the rendering quality and the compactness of the IGS representation. Furthermore, we propose a novel compression pipeline tailored for both point clouds and 2D feature grids, considering the entropy variations across different levels. Extensive experimental evaluations demonstrate that our algorithm can deliver high-quality rendering using only a few MBs, effectively balancing storage efficiency and rendering fidelity, and yielding results that are competitive with the state-of-the-art.

Sergi Masip, Pau Rodriguez, Tinne Tuytelaars et al.

Diffusion models are powerful generative models that achieve state-of-the-art performance in image synthesis. However, training them demands substantial amounts of data and computational resources. Continual learning would allow for incrementally learning new tasks and accumulating knowledge, thus enabling the reuse of trained models for further learning. One potentially suitable continual learning approach is generative replay, where a copy of a generative model trained on previous tasks produces synthetic data that are interleaved with data from the current task. However, standard generative replay applied to diffusion models results in a catastrophic loss in denoising capabilities. In this paper, we propose generative distillation, an approach that distils the entire reverse process of a diffusion model. We demonstrate that our approach substantially improves the continual learning performance of generative replay with only a modest increase in the computational costs.

Vaggelis Dorovatas, Malte Schwerin, Andrew D. Bagdanov et al. · mila

Foundation models have transformed machine learning through large-scale pretraining and increased test-time compute. Despite surpassing human performance in several domains, these models remain fundamentally limited in continuous operation, experience accumulation, and personalization, capabilities that are central to adaptive intelligence. While continual learning research has long targeted these goals, its historical focus on in-weight learning (IWL), i.e., updating a single model's parameters to absorb new knowledge, has rendered catastrophic forgetting a persistent challenge. Our position is that combining the strengths of In-Weight Learning (IWL) and the newly emerged capabilities of In-Context Learning (ICL) through the design of modular memory is the missing piece for continual adaptation at scale. We outline a conceptual framework for modular memory-centric architectures that leverage ICL for rapid adaptation and knowledge accumulation, and IWL for stable updates to model capabilities, charting a practical roadmap toward continually learning agents.

Bo Wan, Yongfei Liu, Desen Zhou et al.

Human object interaction (HOI) detection plays a crucial role in human-centric scene understanding and serves as a fundamental building-block for many vision tasks. One generalizable and scalable strategy for HOI detection is to use weak supervision, learning from image-level annotations only. This is inherently challenging due to ambiguous human-object associations, large search space of detecting HOIs and highly noisy training signal. A promising strategy to address those challenges is to exploit knowledge from large-scale pretrained models (e.g., CLIP), but a direct knowledge distillation strategy~\citep{liao2022gen} does not perform well on the weakly-supervised setting. In contrast, we develop a CLIP-guided HOI representation capable of incorporating the prior knowledge at both image level and HOI instance level, and adopt a self-taught mechanism to prune incorrect human-object associations. Experimental results on HICO-DET and V-COCO show that our method outperforms the previous works by a sizable margin, showing the efficacy of our HOI representation.

Abhishek Jha, Badri N. Patro, Luc Van Gool et al.

Visual question answering is a vision-and-language multimodal task, that aims at predicting answers given samples from the question and image modalities. Most recent methods focus on learning a good joint embedding space of images and questions, either by improving the interaction between these two modalities, or by making it a more discriminant space. However, how informative this joint space is, has not been well explored. In this paper, we propose a novel regularization for VQA models, Constrained Optimization using Barlow's theory (COB), that improves the information content of the joint space by minimizing the redundancy. It reduces the correlation between the learned feature components and thereby disentangles semantic concepts. Our model also aligns the joint space with the answer embedding space, where we consider the answer and image+question as two different `views' of what in essence is the same semantic information. We propose a constrained optimization policy to balance the categorical and redundancy minimization forces. When built on the state-of-the-art GGE model, the resulting model improves VQA accuracy by 1.4% and 4% on the VQA-CP v2 and VQA v2 datasets respectively. The model also exhibits better interpretability.

Timm Hess, Tinne Tuytelaars, Gido M. van de Ven

Recent years have seen considerable progress in the continual training of deep neural networks, predominantly thanks to approaches that add replay or regularization terms to the loss function to approximate the joint loss over all tasks so far. However, we show that even with a perfect approximation to the joint loss, these approaches still suffer from temporary but substantial forgetting when starting to train on a new task. Motivated by this 'stability gap', we propose that continual learning strategies should focus not only on the optimization objective, but also on the way this objective is optimized. While there is some continual learning work that alters the optimization trajectory (e.g., using gradient projection techniques), this line of research is positioned as alternative to improving the optimization objective, while we argue it should be complementary. In search of empirical support for our proposition, we perform a series of pre-registered experiments combining replay-approximated joint objectives with gradient projection-based optimization routines. However, this first experimental attempt fails to show clear and consistent benefits. Nevertheless, our conceptual arguments, as well as some of our empirical results, demonstrate the distinctive importance of the optimization trajectory in continual learning, thereby opening up a new direction for continual learning research.

Manuele Rusci, Tinne Tuytelaars

A personalized KeyWord Spotting (KWS) pipeline typically requires the training of a Deep Learning model on a large set of user-defined speech utterances, preventing fast customization directly applied on-device. To fill this gap, this paper investigates few-shot learning methods for open-set KWS classification by combining a deep feature encoder with a prototype-based classifier. With user-defined keywords from 10 classes of the Google Speech Command dataset, our study reports an accuracy of up to 76% in a 10-shot scenario while the false acceptance rate of unknown data is kept to 5%. In the analyzed settings, the usage of the triplet loss to train an encoder with normalized output features performs better than the prototypical networks jointly trained with a generator of dummy unknown-class prototypes. This design is also more effective than encoders trained on a classification problem and features fewer parameters than other iso-accuracy approaches.

Timm Hess, Eli Verwimp, Gido M. van de Ven et al.

Continual learning research has shown that neural networks suffer from catastrophic forgetting "at the output level", but it is debated whether this is also the case at the level of learned representations. Multiple recent studies ascribe representations a certain level of innate robustness against forgetting -- that they only forget minimally in comparison with forgetting at the output level. We revisit and expand upon the experiments that revealed this difference in forgetting and illustrate the coexistence of two phenomena that affect the quality of continually learned representations: knowledge accumulation and feature forgetting. Taking both aspects into account, we show that, even though forgetting in the representation (i.e. feature forgetting) can be small in absolute terms, when measuring relative to how much was learned during a task, forgetting in the representation tends to be just as catastrophic as forgetting at the output level. Next we show that this feature forgetting is problematic as it substantially slows down the incremental learning of good general representations (i.e. knowledge accumulation). Finally, we study how feature forgetting and knowledge accumulation are affected by different types of continual learning methods.

Bo Wan, Tinne Tuytelaars

In this paper, we investigate the task of zero-shot human-object interaction (HOI) detection, a novel paradigm for identifying HOIs without the need for task-specific annotations. To address this challenging task, we employ CLIP, a large-scale pre-trained vision-language model (VLM), for knowledge distillation on multiple levels. Specifically, we design a multi-branch neural network that leverages CLIP for learning HOI representations at various levels, including global images, local union regions encompassing human-object pairs, and individual instances of humans or objects. To train our model, CLIP is utilized to generate HOI scores for both global images and local union regions that serve as supervision signals. The extensive experiments demonstrate the effectiveness of our novel multi-level CLIP knowledge integration strategy. Notably, the model achieves strong performance, which is even comparable with some fully-supervised and weakly-supervised methods on the public HICO-DET benchmark.

Eli Verwimp, Kuo Yang, Sarah Parisot et al.

Training models continually to detect and classify objects, from new classes and new domains, remains an open problem. In this work, we conduct a thorough analysis of why and how object detection models forget catastrophically. We focus on distillation-based approaches in two-stage networks; the most-common strategy employed in contemporary continual object detection work.Distillation aims to transfer the knowledge of a model trained on previous tasks -- the teacher -- to a new model -- the student -- while it learns the new task. We show that this works well for the region proposal network, but that wrong, yet overly confident teacher predictions prevent student models from effective learning of the classification head. Our analysis provides a foundation that allows us to propose improvements for existing techniques by detecting incorrect teacher predictions, based on current ground-truth labels, and by employing an adaptive Huber loss as opposed to the mean squared error for the distillation loss in the classification heads. We evidence that our strategy works not only in a class incremental setting, but also in domain incremental settings, which constitute a realistic context, likely to be the setting of representative real-world problems.

Manuele Rusci, Francesco Paci, Marco Fariselli et al.

This paper proposes a self-learning method to incrementally train (fine-tune) a personalized Keyword Spotting (KWS) model after the deployment on ultra-low power smart audio sensors. We address the fundamental problem of the absence of labeled training data by assigning pseudo-labels to the new recorded audio frames based on a similarity score with respect to few user recordings. By experimenting with multiple KWS models with a number of parameters up to 0.5M on two public datasets, we show an accuracy improvement of up to +19.2% and +16.0% vs. the initial models pretrained on a large set of generic keywords. The labeling task is demonstrated on a sensor system composed of a low-power microphone and an energy-efficient Microcontroller (MCU). By efficiently exploiting the heterogeneous processing engines of the MCU, the always-on labeling task runs in real-time with an average power cost of up to 8.2 mW. On the same platform, we estimate an energy cost for on-device training 10x lower than the labeling energy if sampling a new utterance every 6.1 s or 18.8 s with a DS-CNN-S or a DS-CNN-M model. Our empirical result paves the way to self-adaptive personalized KWS sensors at the extreme edge.

Vladimir Araujo, Marie-Francine Moens, Tinne Tuytelaars

Parameter-efficient fine-tuning (PEFT) methods are increasingly used with pre-trained language models (PLMs) for continual learning (CL). These methods typically involve training a PEFT module for each new task and employing similarity-based selection to route modules during inference. However, they face two major limitations: 1) interference during module training with already learned modules and 2) suboptimal routing when composing modules. In this paper, we present L2R, a method that isolates the training of new PEFT modules to ensure their task specialization. L2R then learns to compose the learned modules by training a network of routers that leverages a small memory containing examples of previously seen tasks. We evaluate our method in two CL setups using various benchmarks. Our results demonstrate that L2R provides an effective composition of PEFT modules, leading to improved generalization and performance compared to other methods.

Tingyu Qu, Tinne Tuytelaars, Marie-Francine Moens

We revisit the weakly supervised cross-modal face-name alignment task; that is, given an image and a caption, we label the faces in the image with the names occurring in the caption. Whereas past approaches have learned the latent alignment between names and faces by uncertainty reasoning over a set of images and their respective captions, in this paper, we rely on appropriate loss functions to learn the alignments in a neural network setting and propose SECLA and SECLA-B. SECLA is a Symmetry-Enhanced Contrastive Learning-based Alignment model that can effectively maximize the similarity scores between corresponding faces and names in a weakly supervised fashion. A variation of the model, SECLA-B, learns to align names and faces as humans do, that is, learning from easy to hard cases to further increase the performance of SECLA. More specifically, SECLA-B applies a two-stage learning framework: (1) Training the model on an easy subset with a few names and faces in each image-caption pair. (2) Leveraging the known pairs of names and faces from the easy cases using a bootstrapping strategy with additional loss to prevent forgetting and learning new alignments at the same time. We achieve state-of-the-art results for both the augmented Labeled Faces in the Wild dataset and the Celebrity Together dataset. In addition, we believe that our methods can be adapted to other multimodal news understanding tasks.

Zehao Wang, Mingxiao Li, Minye Wu et al.

In this paper, we introduce the map-language navigation task where an agent executes natural language instructions and moves to the target position based only on a given 3D semantic map. To tackle the task, we design the instruction-aware Path Proposal and Discrimination model (iPPD). Our approach leverages map information to provide instruction-aware path proposals, i.e., it selects all potential instruction-aligned candidate paths to reduce the solution space. Next, to represent the map observations along a path for a better modality alignment, a novel Path Feature Encoding scheme tailored for semantic maps is proposed. An attention-based Language Driven Discriminator is designed to evaluate path candidates and determine the best path as the final result. Our method can naturally avoid error accumulation compared with single-step greedy decision methods. Comparing to a single-step imitation learning approach, iPPD has performance gains above 17% on navigation success and 0.18 on path matching measurement nDTW in challenging unseen environments.

Jonathan Swinnen, Tinne Tuytelaars

In this work we introduce a novel approach to domain incremental learning, adapting models over time to evolving, non-stationary data. In contrast to other works, we do not attempt to avoid catastrophic forgetting, but rather allow it and exploit it. Our model combines a main task head with a self-supervised masked autoencoder (MAE) head. We then learn domain-specific LoRA adapters during incremental training. Each adapter specializes to its domain, naturally inducing forgetting on other domains in both heads. At inference, we perform online test-time training on the self-supervised MAE head to identify which LoRAs best matches the current input, so the model can `remember' the domain again. Our scheme is especially well-suited to real-world streaming data, such as video, where consecutive samples are highly correlated and domain shifts are gradual. We demonstrate our method on domain-incremental action recognition and semantic segmentation tasks.

Baptiste Debes, Tinne Tuytelaars

Distributional reinforcement learning (DRL) models the full return distribution rather than expectations, but extending it to multivariate settings remains challenging. Many common metrics do not naturally generalize beyond one dimension or lose computational tractability, and the multivariate case introduces additional difficulties such as general matrix discounting, for which no contraction results are available. We introduce Sliced Distributional Reinforcement Learning (SDRL), which lifts tractable one-dimensional divergences to multivariate return distributions via projections. We prove Bellman contraction for uniform slicing under shared scalar discounting, and introduce a maximum-slicing variant with contraction under general dense discount matrices. SDRL supports a broad class of base divergences; we analyze Wasserstein, Cramér, and Maximum Mean Discrepancy (MMD), and characterize which SDRL variants suit the standard single-sample Bellman update used in distributional RL. We evaluate SDRL on a toy chain problem and a gridworld image-based environment as well as a subset of Atari games.

Simone Marullo, Matteo Tiezzi, Marco Gori et al.

The intrinsic difficulty in adapting deep learning models to non-stationary environments limits the applicability of neural networks to real-world tasks. This issue is critical in practical supervised learning settings, such as the ones in which a pre-trained model computes projections toward a latent space where different task predictors are sequentially learned over time. As a matter of fact, incrementally fine-tuning the whole model to better adapt to new tasks usually results in catastrophic forgetting, with decreasing performance over the past experiences and losing valuable knowledge from the pre-training stage. In this paper, we propose a novel strategy to make the fine-tuning procedure more effective, by avoiding to update the pre-trained part of the network and learning not only the usual classification head, but also a set of newly-introduced learnable parameters that are responsible for transforming the input data. This process allows the network to effectively leverage the pre-training knowledge and find a good trade-off between plasticity and stability with modest computational efforts, thus especially suitable for on-the-edge settings. Our experiments on four image classification problems in a continual learning setting confirm the quality of the proposed approach when compared to several fine-tuning procedures and to popular continual learning methods.

Gorjan Radevski, Dusan Grujicic, Matthew Blaschko et al.

The focal point of egocentric video understanding is modelling hand-object interactions. Standard models -- CNNs, Vision Transformers, etc. -- which receive RGB frames as input perform well, however, their performance improves further by employing additional modalities such as object detections, optical flow, audio, etc. as input. The added complexity of the required modality-specific modules, on the other hand, makes these models impractical for deployment. The goal of this work is to retain the performance of such multimodal approaches, while using only the RGB images as input at inference time. Our approach is based on multimodal knowledge distillation, featuring a multimodal teacher (in the current experiments trained only using object detections, optical flow and RGB frames) and a unimodal student (using only RGB frames as input). We present preliminary results which demonstrate that the resulting model -- distilled from a multimodal teacher -- significantly outperforms the baseline RGB model (trained without knowledge distillation), as well as an omnivorous version of itself (trained on all modalities jointly), in both standard and compositional action recognition.

Abhishek Jha, Tinne Tuytelaars

In this paper, we present an empirical study of typical spatial augmentation techniques used in self-supervised representation learning methods (both contrastive and non-contrastive), namely random crop and cutout. Our contributions are: (a) we dissociate random cropping into two separate augmentations, overlap and patch, and provide a detailed analysis on the effect of area of overlap and patch size to the accuracy on down stream tasks. (b) We offer an insight into why cutout augmentation does not learn good representation, as reported in earlier literature. Finally, based on these analysis, (c) we propose a distance-based margin to the invariance loss for learning scene-centric representations for the downstream task on object-centric distribution, showing that as simple as a margin proportional to the pixel distance between the two spatial views in the scence-centric images can improve the learned representation. Our study furthers the understanding of the spatial augmentations, and the effect of the domain-gap between the training augmentations and the test distribution.

Dipam Goswami, Simone Magistri, Gido M. van de Ven et al.

Vision-language models (VLMs) like CLIP are trained with the objective of aligning text and image pairs. To improve CLIP-based few-shot image classification, recent works have observed that, along with text embeddings, image embeddings from the training set are an important source of information. In this work we investigate the impact of directly mixing image and text prototypes for few-shot classification and analyze this from a bias-variance perspective. We show that mixing prototypes acts like a shrinkage estimator. Although mixed prototypes improve classification performance, the image prototypes still add some noise in the form of instance-specific background or context information. In order to capture only information from the image space relevant to the given classification task, we propose projecting image prototypes onto the principal directions of the semantic text embedding space to obtain a text-aligned semantic image subspace. These text-aligned image prototypes, when mixed with text embeddings, further improve classification. However, for downstream datasets with poor cross-modal alignment in CLIP, semantic alignment might be suboptimal. We show that the image subspace can still be leveraged by modeling the anisotropy using class covariances. We demonstrate that combining a text-aligned mixed prototype classifier and an image-specific LDA classifier outperforms existing methods across few-shot classification benchmarks.

Fanis Mathioulakis, Gorjan Radevski, Tinne Tuytelaars

We introduce Eff-GRot, an approach for efficient and generalizable rotation estimation from RGB images. Given a query image and a set of reference images with known orientations, our method directly predicts the object's rotation in a single forward pass, without requiring object- or category-specific training. At the core of our framework is a transformer that performs a comparison in the latent space, jointly processing rotation-aware representations from multiple references alongside a query. This design enables a favorable balance between accuracy and computational efficiency while remaining simple, scalable, and fully end-to-end. Experimental results show that Eff-GRot offers a promising direction toward more efficient rotation estimation, particularly in latency-sensitive applications.

Sergi Masip, Gido M. van de Ven, Javier Ferrando et al.

Catastrophic forgetting in continual learning is often measured at the performance or last-layer representation level, overlooking the underlying mechanisms. We introduce a mechanistic framework that offers a geometric interpretation of catastrophic forgetting as the result of transformations to the encoding of individual features. These transformations can lead to forgetting by reducing the allocated capacity of features (worse representation) and disrupting their readout by downstream computations. Analysis of a tractable model formalizes this view, allowing us to identify best- and worst-case scenarios. Through experiments on this model, we empirically test our formal analysis and highlight the detrimental effect of depth. Finally, we demonstrate how our framework can be used in the analysis of practical models through the use of Crosscoders. We present a case study of a Vision Transformer trained on sequential CIFAR-10. Our work provides a new, feature-centric vocabulary for continual learning.

Pierre-François De Plaen, Abhishek Jha, Luc Van Gool et al.

Self-supervised learning (SSL) methods based on Siamese networks learn visual representations by aligning different views of the same image. The multi-crop strategy, which incorporates small local crops to global ones, enhances many SSL frameworks but causes instability in predictor-based architectures such as BYOL, SimSiam, and MoCo v3. We trace this failure to the shared predictor used across all views and demonstrate that assigning a separate predictor to each view type stabilizes multi-crop training, resulting in significant performance gains. Extending this idea, we treat each spatial transformation as a distinct alignment task and add cutout views, where part of the image is masked before encoding. This yields a simple multi-task formulation of asymmetric Siamese SSL that combines global, local, and masked views into a single framework. The approach is stable, generally applicable across backbones, and consistently improves the performance of ResNet and ViT models on ImageNet.

Karel Moens, Matthew B. Blaschko, Tinne Tuytelaars et al.

Whole slide image (WSI) normalization remains a vital preprocessing step in computational pathology. Increasingly driven by deep learning, these models learn to approximate data distributions from training examples. This often results in outputs that gravitate toward the average, potentially masking diagnostically important features. More critically, they can introduce hallucinated content, artifacts that appear realistic but are not present in the original tissue, posing a serious threat to downstream analysis. These hallucinations are nearly impossible to detect visually, and current evaluation practices often overlook them. In this work, we demonstrate that the risk of hallucinations is real and underappreciated. While many methods perform adequately on public datasets, we observe a concerning frequency of hallucinations when these same models are retrained and evaluated on real-world clinical data. To address this, we propose a novel image comparison measure designed to automatically detect hallucinations in normalized outputs. Using this measure, we systematically evaluate several well-cited normalization methods retrained on real-world data, revealing significant inconsistencies and failures that are not captured by conventional metrics. Our findings underscore the need for more robust, interpretable normalization techniques and stricter validation protocols in clinical deployment.

Cédric Picron, Tinne Tuytelaars

Many two-stage instance segmentation heads predict a coarse 28x28 mask per instance, which is insufficient to capture the fine-grained details of many objects. To address this issue, PointRend and RefineMask predict a 112x112 segmentation mask resulting in higher quality segmentations. Both methods however have limitations by either not having access to neighboring features (PointRend) or by performing computation at all spatial locations instead of sparsely (RefineMask). In this work, we propose EffSeg performing fine-grained instance segmentation in an efficient way by using our Structure-Preserving Sparsity (SPS) method based on separately storing the active features, the passive features and a dense 2D index map containing the feature indices. The goal of the index map is to preserve the 2D spatial configuration or structure between the features such that any 2D operation can still be performed. EffSeg achieves similar performance on COCO compared to RefineMask, while reducing the number of FLOPs by 71% and increasing the FPS by 29%. Code will be released.

Jonathan Tonglet, Tinne Tuytelaars, Marie-Francine Moens et al.

Visualizations play a pivotal role in daily communication in an increasingly datadriven world. Research on multimodal large language models (MLLMs) for automated chart understanding has accelerated massively, with steady improvements on standard benchmarks. However, for MLLMs to be reliable, they must be robust to misleading visualizations, i.e., charts that distort the underlying data, leading readers to draw inaccurate conclusions that may support disinformation. Here, we uncover an important vulnerability: MLLM questionanswering (QA) accuracy on misleading visualizations drops on average to the level of the random baseline. To address this, we introduce the first inference-time methods to improve QA performance on misleading visualizations, without compromising accuracy on non-misleading ones. We find that two methods, table-based QA and redrawing the visualization, are effective, with improvements of up to 19.6 percentage points. We make our code and data available.