Compositional generalization, the ability to generate novel combinations of known concepts, is a key ingredient for visual generative models. Yet, not all mechanisms that enable or inhibit it are fully understood. In this work, we conduct a systematic study of how various design choices influence compositional generalization in image and video generation in a positive or negative way. Through controlled experiments, we identify two key factors: (i) whether the training objective operates on a discrete or continuous distribution, and (ii) to what extent conditioning provides information about the constituent concepts during training. Building on these insights, we show that relaxing the MaskGIT discrete loss with an auxiliary continuous JEPA-based objective can improve compositional performance in discrete models like MaskGIT.

Modern challenges of robustness, fairness, and decision-making in machine learning have led to the formulation of multi-distribution learning (MDL) frameworks in which a predictor is optimized across multiple distributions. We study the calibration properties of MDL to better understand how the predictor performs uniformly across the multiple distributions. Through classical results on decomposing proper scoring losses, we first derive the Bayes optimal rule for MDL, demonstrating that it maximizes the generalized entropy of the associated loss function. Our analysis reveals that while this approach ensures minimal worst-case loss, it can lead to non-uniform calibration errors across the multiple distributions and there is an inherent calibration-refinement trade-off, even at Bayes optimality. Our results highlight a critical limitation: despite the promise of MDL, one must use caution when designing predictors tailored to multiple distributions so as to minimize disparity.

Contrastive vision-language models (VLMs), like CLIP, have gained popularity for their versatile applicability to various downstream tasks. Despite their successes in some tasks, like zero-shot object recognition, they perform surprisingly poor on other tasks, like attribute recognition. Previous work has attributed these challenges to the modality gap, a separation of image and text in the shared representation space, and to a bias towards objects over other factors, such as attributes. In this analysis paper, we investigate both phenomena thoroughly. We evaluated off-the-shelf VLMs and while the gap's influence on performance is typically overshadowed by other factors, we find indications that closing the gap indeed leads to improvements. Moreover, we find that, contrary to intuition, only few embedding dimensions drive the gap and that the embedding spaces are differently organized. To allow for a clean study of object bias, we introduce a definition and a corresponding measure of it. Equipped with this tool, we find that object bias does not lead to worse performance on other concepts, such as attributes per se. However, why do both phenomena, modality gap and object bias, emerge in the first place? To answer this fundamental question and uncover some of the inner workings of contrastive VLMs, we conducted experiments that allowed us to control the amount of shared information between the modalities. These experiments revealed that the driving factor behind both the modality gap and the object bias, is an information imbalance between images and captions, and unveiled an intriguing connection between the modality gap and entropy of the logits.

Pretrained vision-language models, such as CLIP, show promising zero-shot performance across a wide variety of datasets. For closed-set classification tasks, however, there is an inherent limitation: CLIP image encoders are typically designed to extract generic image-level features that summarize superfluous or confounding information for the target tasks. This results in degradation of classification performance, especially when objects of interest cover small areas of input images. In this work, we propose CLIP with Guided Cropping (GC-CLIP), where we use an off-the-shelf zero-shot object detection model in a preprocessing step to increase focus of zero-shot classifier to the object of interest and minimize influence of extraneous image regions. We empirically show that our approach improves zero-shot classification results across architectures and datasets, favorably for small objects.

In this work, we study rapid improvements of the training loss in transformers when being confronted with multi-step decision tasks. We found that transformers struggle to learn the intermediate task and both training and validation loss saturate for hundreds of epochs. When transformers finally learn the intermediate task, they do this rapidly and unexpectedly. We call these abrupt improvements Eureka-moments, since the transformer appears to suddenly learn a previously incomprehensible concept. We designed synthetic tasks to study the problem in detail, but the leaps in performance can be observed also for language modeling and in-context learning (ICL). We suspect that these abrupt transitions are caused by the multi-step nature of these tasks. Indeed, we find connections and show that ways to improve on the synthetic multi-step tasks can be used to improve the training of language modeling and ICL. Using the synthetic data we trace the problem back to the Softmax function in the self-attention block of transformers and show ways to alleviate the problem. These fixes reduce the required number of training steps, lead to higher likelihood to learn the intermediate task, to higher final accuracy and training becomes more robust to hyper-parameters.

Shortcut learning occurs when a deep neural network overly relies on spurious correlations in the training dataset in order to solve downstream tasks. Prior works have shown how this impairs the compositional generalization capability of deep learning models. To address this problem, we propose a novel approach to mitigate shortcut learning in uncontrolled target domains. Our approach extends the training set with an additional dataset (the source domain), which is specifically designed to facilitate learning independent representations of basic visual factors. We benchmark our idea on synthetic target domains where we explicitly control shortcut opportunities as well as real-world target domains. Furthermore, we analyze the effect of different specifications of the source domain and the network architecture on compositional generalization. Our main finding is that leveraging data from a source domain is an effective way to mitigate shortcut learning. By promoting independence across different factors of variation in the learned representations, networks can learn to consider only predictive factors and ignore potential shortcut factors during inference.

The standard approach to contrastive learning is to maximize the agreement between different views of the data. The views are ordered in pairs, such that they are either positive, encoding different views of the same object, or negative, corresponding to views of different objects. The supervisory signal comes from maximizing the total similarity over positive pairs, while the negative pairs are needed to avoid collapse. In this work, we note that the approach of considering individual pairs cannot account for both intra-set and inter-set similarities when the sets are formed from the views of the data. It thus limits the information content of the supervisory signal available to train representations. We propose to go beyond contrasting individual pairs of objects by focusing on contrasting objects as sets. For this, we use combinatorial quadratic assignment theory designed to evaluate set and graph similarities and derive set-contrastive objective as a regularizer for contrastive learning methods. We conduct experiments and demonstrate that our method improves learned representations for the tasks of metric learning and self-supervised classification.

Open Set Recognition (OSR) extends image classification to an open-world setting, by simultaneously classifying known classes and identifying unknown ones. While conventional OSR approaches can detect Out-of-Distribution (OOD) samples, they cannot provide explanations indicating which underlying visual attribute(s) (e.g., shape, color or background) cause a specific sample to be unknown. In this work, we introduce a novel problem setup that generalizes conventional OSR to a multi-attribute setting, where multiple visual attributes are simultaneously recognized. Here, OOD samples can be not only identified but also categorized by their unknown attribute(s). We propose simple extensions of common OSR baselines to handle this novel scenario. We show that these baselines are vulnerable to shortcuts when spurious correlations exist in the training dataset. This leads to poor OOD performance which, according to our experiments, is mainly due to unintended cross-attribute correlations of the predicted confidence scores. We provide an empirical evidence showing that this behavior is consistent across different baselines on both synthetic and real world datasets.

Speech Activity Detection (SAD), locating speech segments within an audio recording, is a main part of most speech technology applications. Robust SAD is usually more difficult in noisy conditions with varying signal-to-noise ratios (SNR). The Fearless Steps challenge has recently provided such data from the NASA Apollo-11 mission for different speech processing tasks including SAD. Most audio recordings are degraded by different kinds and levels of noise varying within and between channels. This paper describes the EML online algorithm for the most recent phase of this challenge. The proposed algorithm can be trained both in a supervised and unsupervised manner and assigns speech and non-speech labels at runtime approximately every 0.1 sec. The experimental results show a competitive accuracy on both development and evaluation datasets with a real-time factor of about 0.002 using a single CPU machine.

Convolutional neural networks (CNNs) learn to extract representations of complex features, such as object shapes and textures to solve image recognition tasks. Recent work indicates that CNNs trained on ImageNet are biased towards features that encode textures and that these alone are sufficient to generalize to unseen test data from the same distribution as the training data but often fail to generalize to out-of-distribution data. It has been shown that augmenting the training data with different image styles decreases this texture bias in favor of increased shape bias while at the same time improving robustness to common corruptions, such as noise and blur. Commonly, this is interpreted as shape bias increasing corruption robustness. However, this relationship is only hypothesized. We perform a systematic study of different ways of composing inputs based on natural images, explicit edge information, and stylization. While stylization is essential for achieving high corruption robustness, we do not find a clear correlation between shape bias and robustness. We conclude that the data augmentation caused by style-variation accounts for the improved corruption robustness and increased shape bias is only a byproduct.

Common deep neural networks (DNNs) for image classification have been shown to rely on shortcut opportunities (SO) in the form of predictive and easy-to-represent visual factors. This is known as shortcut learning and leads to impaired generalization. In this work, we show that common DNNs also suffer from shortcut learning when predicting only basic visual object factors of variation (FoV) such as shape, color, or texture. We argue that besides shortcut opportunities, generalization opportunities (GO) are also an inherent part of real-world vision data and arise from partial independence between predicted classes and FoVs. We also argue that it is necessary for DNNs to exploit GO to overcome shortcut learning. Our core contribution is to introduce the Diagnostic Vision Benchmark suite DiagViB-6, which includes datasets and metrics to study a network's shortcut vulnerability and generalization capability for six independent FoV. In particular, DiagViB-6 allows controlling the type and degree of SO and GO in a dataset. We benchmark a wide range of popular vision architectures and show that they can exploit GO only to a limited extent.

Single-view 3D object reconstruction has seen much progress, yet methods still struggle generalizing to novel shapes unseen during training. Common approaches predominantly rely on learned global shape priors and, hence, disregard detailed local observations. In this work, we address this issue by learning a hierarchy of priors at different levels of locality from ground truth input depth maps. We argue that exploiting local priors allows our method to efficiently use input observations, thus improving generalization in visible areas of novel shapes. At the same time, the combination of local and global priors enables meaningful hallucination of unobserved parts resulting in consistent 3D shapes. We show that the hierarchical approach generalizes much better than the global approach. It generalizes not only between different instances of a class but also across classes and to unseen arrangements of objects.

We introduce the SE(3)-Transformer, a variant of the self-attention module for 3D point clouds, which is equivariant}under continuous 3D roto-translations. Equivariance is important to ensure stable and predictable performance in the presence of nuisance transformations of the data input. A positive corollary of equivariance is increased weight-tying within the model, leading to fewer trainable parameters and thus decreased sample complexity (i.e. we need less training data). The SE(3)-Transformer leverages the benefits of self-attention to operate on large point clouds with varying number of points, while guaranteeing SE(3)-equivariance for robustness. We evaluate our model on a toy $N$-body particle simulation dataset, showcasing the robustness of the predictions under rotations of the input. We further achieve competitive performance on two real-world datasets, ScanObjectNN and QM9. In all cases, our model outperforms a strong, non-equivariant attention baseline and an equivariant model without attention.

This technical report describes the EML submission to the first VoxCeleb speaker diarization challenge. Although the aim of the challenge has been the offline processing of the signals, the submitted system is basically the EML online algorithm which decides about the speaker labels in runtime approximately every 1.2 sec. For the first phase of the challenge, only VoxCeleb2 dev dataset was used for training. The results on the provided VoxConverse dev set show much better accuracy in terms of both DER and JER compared to the offline baseline provided in the challenge. The real-time factor of the whole diarization process is about 0.01 using a single CPU machine.

Recently, there has been a growing interest in developing saliency methods that provide visual explanations of network predictions. Still, the usability of existing methods is limited to image classification models. To overcome this limitation, we extend the existing approaches to generate grid saliencies, which provide spatially coherent visual explanations for (pixel-level) dense prediction networks. As the proposed grid saliency allows to spatially disentangle the object and its context, we specifically explore its potential to produce context explanations for semantic segmentation networks, discovering which context most influences the class predictions inside a target object area. We investigate the effectiveness of grid saliency on a synthetic dataset with an artificially induced bias between objects and their context as well as on the real-world Cityscapes dataset using state-of-the-art segmentation networks. Our results show that grid saliency can be successfully used to provide easily interpretable context explanations and, moreover, can be employed for detecting and localizing contextual biases present in the data.

In this paper, we report some recent improvements to DNN/HMM hybrid acoustic modeling for the EML real-time large vocabulary speech recognition system, including the introduction of speaker adaptive long short term memory units (LSTMs) and efficient online decoding with deep bidirectional LSTMs. Based on a thorough latency analysis of our baseline large vocabulary speech recognizer we first abandon multi-pass recognition with fMLLR adapted acoustic features and further simplify decoding by dropping text independent vocal tract length normalization (VTLN) which was identified as a major bottleneck for real time applications. Subsequently, we improve accuracy by a variety of measures that include artificial training data augmentation and the use of additional features derived from an online speaker diarization module currently under development. Moreover, we investigate a hierarchy of feed forward and recurrent neural networks for a further reduction of word error rate. Finally, we demonstrate that established DNN pruning techniques are also applicable to bidirectional LSTMs, resulting in both an appropriate network size and substantial runtime savings. Our experiments are carried out on the publicly available WSJCAM0 corpus. Being simultaneously recorded with both a headmounted and a desk-mounted microphone it enables us to study the impact of each of the proposed methods also in case of a channel mismatch between training and test data. The methods described in this paper yield an improvement of up to 15 percent relative to the baseline DNN/HMM acoustic model.

Speaker diarization is more challenging in presence of background noise or music, frequent speaker changes, and cross talks. In an online scenario, the decision should be made at time, given only the current short segment and the speakers detected in the past, which makes the task even harder. In this work, an online robust speaker diarization algorithm is proposed in which speech segments are represented by low dimensional vectors referred to as speaker-corrupted embeddings. The proposed speaker embedding network is a deep neural network which takes speaker-corrupted supervectors as input, uses variable ReLU (VReLU) as an activation function, and tries to discriminate the background speakers. Speaker corruption is performed by adding supervectors built by 20 speech frames from other speakers to the supervectors of a given speaker. It is shown that speaker corruption, VReLU, and input dropout increase the generalization power of the proposed network. To increase the robustness, the proposed embeddings are concatenated with LDA transformed supervectors. Experimental results on the Albayzin 2018 evaluation set show a competitive accuracy, more robustness, and much lower computational cost compared to typical offline algorithms.

An environment representation (ER) is a substantial part of every autonomous system. It introduces a common interface between perception and other system components, such as decision making, and allows downstream algorithms to deal with abstract data without knowledge of the used sensor. In this work, we propose and evaluate a novel architecture that generates an egocentric, grid-based, predictive, and semantically-interpretable ER, which we call semantic grid. We show that our approach supports the spatio-temporal fusion of multiple camera sequences and short-term prediction in such an ER. Our design utilizes a strong semantic segmentation network together with depth and egomotion estimates to first extract semantic information from multiple camera streams and then transform these separately into egocentric temporally-aligned bird's-eye view grids. A deep encoder-decoder network is trained to fuse a stack of these grids into a unified semantic grid and to predict the dynamics of its surrounding. We evaluate this representation on real-world sequences of Cityscapes and show that our architecture can make accurate predictions in complex sensor fusion scenarios and significantly outperforms a model-driven baseline in a category-based evaluation.

Deep neural networks achieve state-of-the-art results on several tasks while increasing in complexity. It has been shown that neural networks can be pruned during training by imposing sparsity inducing regularizers. In this paper, we investigate two techniques for group-wise pruning during training in order to improve network efficiency. We propose a gating factor after every convolutional layer to induce channel level sparsity, encouraging insignificant channels to become exactly zero. Further, we introduce and analyse a bounded variant of the ℓ _1 regularizer, which interpolates between ℓ _1 and ℓ _0 -norms to retain performance of the network at higher pruning rates. To underline effectiveness of the proposed methods, we show that the number of parameters of ResNet-164, DenseNet-40 and MobileNetV2 can be reduced down by 30% , 69% , and 75% on CIFAR100 respectively without a significant drop in accuracy. We achieve state-of-the-art pruning results for ResNet-50 with higher accuracy on ImageNet. Furthermore, we show that the light weight MobileNetV2 can further be compressed on ImageNet without a significant drop in performance .

Voice Activity Detection (VAD), locating speech segments within an au- dio recording, is a main part of most speech technology applications. Non-speech segments, e.g., silence, noise, and music, usually do not carry any interesting infor- mation in speech recognition applications and they even degrade the performance of the recognition system in terms of both the accuracy and computational cost. Various VAD techniques have been developed, but not all of them are appropriate for a real-time application where the robustness, accuracy, and the processing time are the main keys. In this paper, we propose a fast and robust VAD for a real-time Automatic Speech Recognition (ASR) task. The main goal is to efficiently filter out the non-speech segments before processing the speech segments of the audio signal by the decoder. The proposed technique is a hybrid supervised/unsupervised model based on zero-order Baum-Welch statistics obtained from a Universal Back- ground Model (UBM).We will show that not only the processing time for the whole speech recognition task is decreased by 39%, but also the Word Error Rate (WER) is reduced by about 1.9% relative.