In most E-commerce platforms, whether the displayed items trigger the user's interest largely depends on their most eye-catching multimodal content. Consequently, increasing efforts focus on modeling multimodal user preference, and the pressing paradigm is to incorporate complete multimodal deep features of the items into the recommendation module. However, the existing studies ignore the mismatch problem between multimodal feature extraction (MFE) and user interest modeling (UIM). That is, MFE and UIM have different emphases. Specifically, MFE is migrated from and adapted to upstream tasks such as image classification. In addition, it is mainly a content-oriented and non-personalized process, while UIM, with its greater focus on understanding user interaction, is essentially a user-oriented and personalized process. Therefore, the direct incorporation of MFE into UIM for purely user-oriented tasks, tends to introduce a large number of preference-independent multimodal noise and contaminate the embedding representations in UIM. This paper aims at solving the mismatch problem between MFE and UIM, so as to generate high-quality embedding representations and better model multimodal user preferences. Towards this end, we develop a novel model, MEGCF. The UIM of the proposed model captures the semantic correlation between interactions and the features obtained from MFE, thus making a better match between MFE and UIM. More precisely, semantic-rich entities are first extracted from the multimodal data, since they are more relevant to user preferences than other multimodal information. These entities are then integrated into the user-item interaction graph. Afterwards, a symmetric linear Graph Convolution Network (GCN) module is constructed to perform message propagation over the graph, in order to capture both high-order semantic correlation and collaborative filtering signals.
Mixup is a commonly adopted data augmentation technique for image classification. Recent advances in mixup methods primarily focus on mixing based on saliency. However, many saliency detectors require intense computation and are especially burdensome for parameter-heavy transformer models. To this end, we propose TokenMixup, an efficient attention-guided token-level data augmentation method that aims to maximize the saliency of a mixed set of tokens. TokenMixup provides x15 faster saliency-aware data augmentation compared to gradient-based methods. Moreover, we introduce a variant of TokenMixup which mixes tokens within a single instance, thereby enabling multi-scale feature augmentation. Experiments show that our methods significantly improve the baseline models' performance on CIFAR and ImageNet-1K, while being more efficient than previous methods. We also reach state-of-the-art performance on CIFAR-100 among from-scratch transformer models. Code is available at https://github.com/mlvlab/TokenMixup.
We propose learnable polyphase sampling (LPS), a pair of learnable down/upsampling layers that enable truly shift-invariant and equivariant convolutional networks. LPS can be trained end-to-end from data and generalizes existing handcrafted downsampling layers. It is widely applicable as it can be integrated into any convolutional network by replacing down/upsampling layers. We evaluate LPS on image classification and semantic segmentation. Experiments show that LPS is on-par with or outperforms existing methods in both performance and shift consistency. For the first time, we achieve true shift-equivariance on semantic segmentation (PASCAL VOC), i.e., 100% shift consistency, outperforming baselines by an absolute 3.3%.
Data augmentation is key to improving the generalization ability of deep learning models. Mixup is a simple and widely-used data augmentation technique that has proven effective in alleviating the problems of overfitting and data scarcity. Also, recent studies of saliency-aware Mixup in the image domain show that preserving discriminative parts is beneficial to improving the generalization performance. However, these Mixup-based data augmentations are underexplored in 3D vision, especially in point clouds. In this paper, we propose SageMix, a saliency-guided Mixup for point clouds to preserve salient local structures. Specifically, we extract salient regions from two point clouds and smoothly combine them into one continuous shape. With a simple sequential sampling by re-weighted saliency scores, SageMix preserves the local structure of salient regions. Extensive experiments demonstrate that the proposed method consistently outperforms existing Mixup methods in various benchmark point cloud datasets. With PointNet++, our method achieves an accuracy gain of 2.6% and 4.0% over standard training in 3D Warehouse dataset (MN40) and ScanObjectNN, respectively. In addition to generalization performance, SageMix improves robustness and uncertainty calibration. Moreover, when adopting our method to various tasks including part segmentation and standard 2D image classification, our method achieves competitive performance.
A Bayesian pseudocoreset is a small synthetic dataset for which the posterior over parameters approximates that of the original dataset. While promising, the scalability of Bayesian pseudocoresets is not yet validated in realistic problems such as image classification with deep neural networks. On the other hand, dataset distillation methods similarly construct a small dataset such that the optimization using the synthetic dataset converges to a solution with performance competitive with optimization using full data. Although dataset distillation has been empirically verified in large-scale settings, the framework is restricted to point estimates, and their adaptation to Bayesian inference has not been explored. This paper casts two representative dataset distillation algorithms as approximations to methods for constructing pseudocoresets by minimizing specific divergence measures: reverse KL divergence and Wasserstein distance. Furthermore, we provide a unifying view of such divergence measures in Bayesian pseudocoreset construction. Finally, we propose a novel Bayesian pseudocoreset algorithm based on minimizing forward KL divergence. Our empirical results demonstrate that the pseudocoresets constructed from these methods reflect the true posterior even in high-dimensional Bayesian inference problems.
Neural networks are known to produce poor uncertainty estimations, and a variety of approaches have been proposed to remedy this issue. This includes deep ensemble, a simple and effective method that achieves state-of-the-art results for uncertainty-aware learning tasks. In this work, we explore a combinatorial generalization of deep ensemble called deep combinatorial aggregation (DCA). DCA creates multiple instances of network components and aggregates their combinations to produce diversified model proposals and predictions. DCA components can be defined at different levels of granularity. And we discovered that coarse-grain DCAs can outperform deep ensemble for uncertainty-aware learning both in terms of predictive performance and uncertainty estimation. For fine-grain DCAs, we discover that an average parameterization approach named deep combinatorial weight averaging (DCWA) can improve the baseline training. It is on par with stochastic weight averaging (SWA) but does not require any custom training schedule or adaptation of BatchNorm layers. Furthermore, we propose a consistency enforcing loss that helps the training of DCWA and modelwise DCA. We experiment on in-domain, distributional shift, and out-of-distribution image classification tasks, and empirically confirm the effectiveness of DCWA and DCA approaches.
Few-shot learning (FSL) has attracted considerable attention recently. Among existing approaches, the metric-based method aims to train an embedding network that can make similar samples close while dissimilar samples as far as possible and achieves promising results. FSL is characterized by using only a few images to train a model that can generalize to novel classes in image classification problems, but this setting makes it difficult to learn the visual features that can identify the images' appearance variations. The model training is likely to move in the wrong direction, as the images in an identical semantic class may have dissimilar appearances, whereas the images in different semantic classes may share a similar appearance. We argue that FSL can benefit from additional semantic features to learn discriminative feature representations. Thus, this study proposes a multi-task learning approach to view semantic features of label text as an auxiliary task to help boost the performance of the FSL task. Our proposed model uses word-embedding representations as semantic features to help train the embedding network and a semantic cross-attention module to bridge the semantic features into the typical visual modal. The proposed approach is simple, but produces excellent results. We apply our proposed approach to two previous metric-based FSL methods, all of which can substantially improve performance. The source code for our model is accessible from github.
Spatial-wise dynamic convolution has become a promising approach to improving the inference efficiency of deep networks. By allocating more computation to the most informative pixels, such an adaptive inference paradigm reduces the spatial redundancy in image features and saves a considerable amount of unnecessary computation. However, the theoretical efficiency achieved by previous methods can hardly translate into a realistic speedup, especially on the multi-core processors (e.g. GPUs). The key challenge is that the existing literature has only focused on designing algorithms with minimal computation, ignoring the fact that the practical latency can also be influenced by scheduling strategies and hardware properties. To bridge the gap between theoretical computation and practical efficiency, we propose a latency-aware spatial-wise dynamic network (LASNet), which performs coarse-grained spatially adaptive inference under the guidance of a novel latency prediction model. The latency prediction model can efficiently estimate the inference latency of dynamic networks by simultaneously considering algorithms, scheduling strategies, and hardware properties. We use the latency predictor to guide both the algorithm design and the scheduling optimization on various hardware platforms. Experiments on image classification, object detection and instance segmentation demonstrate that the proposed framework significantly improves the practical inference efficiency of deep networks. For example, the average latency of a ResNet-101 on the ImageNet validation set could be reduced by 36% and 46% on a server GPU (Nvidia Tesla-V100) and an edge device (Nvidia Jetson TX2 GPU) respectively without sacrificing the accuracy. Code is available at https://github.com/LeapLabTHU/LASNet.
Recent years have witnessed the rise and success of pre-training techniques in visually-rich document understanding. However, most existing methods lack the systematic mining and utilization of layout-centered knowledge, leading to sub-optimal performances. In this paper, we propose ERNIE-Layout, a novel document pre-training solution with layout knowledge enhancement in the whole workflow, to learn better representations that combine the features from text, layout, and image. Specifically, we first rearrange input sequences in the serialization stage, and then present a correlative pre-training task, reading order prediction, to learn the proper reading order of documents. To improve the layout awareness of the model, we integrate a spatial-aware disentangled attention into the multi-modal transformer and a replaced regions prediction task into the pre-training phase. Experimental results show that ERNIE-Layout achieves superior performance on various downstream tasks, setting new state-of-the-art on key information extraction, document image classification, and document question answering datasets. The code and models are publicly available at http://github.com/PaddlePaddle/PaddleNLP/tree/develop/model_zoo/ernie-layout.
Learning medical visual representations directly from paired radiology reports has become an emerging topic in representation learning. However, existing medical image-text joint learning methods are limited by instance or local supervision analysis, ignoring disease-level semantic correspondences. In this paper, we present a novel Multi-Granularity Cross-modal Alignment (MGCA) framework for generalized medical visual representation learning by harnessing the naturally exhibited semantic correspondences between medical image and radiology reports at three different levels, i.e., pathological region-level, instance-level, and disease-level. Specifically, we first incorporate the instance-wise alignment module by maximizing the agreement between image-report pairs. Further, for token-wise alignment, we introduce a bidirectional cross-attention strategy to explicitly learn the matching between fine-grained visual tokens and text tokens, followed by contrastive learning to align them. More important, to leverage the high-level inter-subject relationship semantic (e.g., disease) correspondences, we design a novel cross-modal disease-level alignment paradigm to enforce the cross-modal cluster assignment consistency. Extensive experimental results on seven downstream medical image datasets covering image classification, object detection, and semantic segmentation tasks demonstrate the stable and superior performance of our framework.
Classification on smartphone-captured chest X-ray (CXR) photos to detect pathologies is challenging due to the projective transformation caused by the non-ideal camera position. Recently, various rectification methods have been proposed for different photo rectification tasks such as document photos, license plate photos, etc. Unfortunately, we found that none of them is suitable for CXR photos, due to their specific transformation type, image appearance, annotation type, etc. In this paper, we propose an innovative deep learning-based Projective Transformation Rectification Network (PTRN) to automatically rectify CXR photos by predicting the projective transformation matrix. To the best of our knowledge, it is the first work to predict the projective transformation matrix as the learning goal for photo rectification. Additionally, to avoid the expensive collection of natural data, synthetic CXR photos are generated under the consideration of natural perturbations, extra screens, etc. We evaluate the proposed approach in the CheXphoto smartphone-captured CXR photos classification competition hosted by the Stanford University Machine Learning Group, our approach won first place with a huge performance improvement (ours 0.850, second-best 0.762, in AUC). A deeper study demonstrates that the use of PTRN successfully achieves the classification performance on the spatially transformed CXR photos to the same level as on the high-quality digital CXR images, indicating PTRN can eliminate all negative impacts of projective transformation on the CXR photos.
Data mixing strategies (e.g., CutMix) have shown the ability to greatly improve the performance of convolutional neural networks (CNNs). They mix two images as inputs for training and assign them with a mixed label with the same ratio. While they are shown effective for vision transformers (ViTs), we identify a token fluctuation phenomenon that has suppressed the potential of data mixing strategies. We empirically observe that the contributions of input tokens fluctuate as forward propagating, which might induce a different mixing ratio in the output tokens. The training target computed by the original data mixing strategy can thus be inaccurate, resulting in less effective training. To address this, we propose a token-label alignment (TL-Align) method to trace the correspondence between transformed tokens and the original tokens to maintain a label for each token. We reuse the computed attention at each layer for efficient token-label alignment, introducing only negligible additional training costs. Extensive experiments demonstrate that our method improves the performance of ViTs on image classification, semantic segmentation, objective detection, and transfer learning tasks. Code is available at: https://github.com/Euphoria16/TL-Align.
Spiking Neural Networks (SNNs), as one of the algorithmic models in neuromorphic computing, have gained a great deal of research attention owing to temporal information processing capability, low power consumption, and high biological plausibility. The potential to efficiently extract spatio-temporal features makes it suitable for processing the event streams. However, existing synaptic structures in SNNs are almost full-connections or spatial 2D convolution, neither of which can extract temporal dependencies adequately. In this work, we take inspiration from biological synapses and propose a spatio-temporal synaptic connection SNN (STSC-SNN) model, to enhance the spatio-temporal receptive fields of synaptic connections, thereby establishing temporal dependencies across layers. Concretely, we incorporate temporal convolution and attention mechanisms to implement synaptic filtering and gating functions. We show that endowing synaptic models with temporal dependencies can improve the performance of SNNs on classification tasks. In addition, we investigate the impact of performance vias varied spatial-temporal receptive fields and reevaluate the temporal modules in SNNs. Our approach is tested on neuromorphic datasets, including DVS128 Gesture (gesture recognition), N-MNIST, CIFAR10-DVS (image classification), and SHD (speech digit recognition). The results show that the proposed model outperforms the state-of-the-art accuracy on nearly all datasets.
A vast literature shows that the learning-based visual perception model is sensitive to adversarial noises, but few works consider the robustness of robotic perception models under widely-existing camera motion perturbations. To this end, we study the robustness of the visual perception model under camera motion perturbations to investigate the influence of camera motion on robotic perception. Specifically, we propose a motion smoothing technique for arbitrary image classification models, whose robustness under camera motion perturbations could be certified. The proposed robustness certification framework based on camera motion smoothing provides tight and scalable robustness guarantees for visual perception modules so that they are applicable to wide robotic applications. As far as we are aware, this is the first work to provide robustness certification for the deep perception module against camera motions, which improves the trustworthiness of robotic perception. A realistic indoor robotic dataset with a dense point cloud map for the entire room, MetaRoom, is introduced for the challenging certifiable robust perception task. We conduct extensive experiments to validate the certification approach via motion smoothing against camera motion perturbations. Our framework guarantees the certified accuracy of 81.7% against camera translation perturbation along depth direction within -0.1m ~ 0.1m. We also validate the effectiveness of our method on the real-world robot by conducting hardware experiments on the robotic arm with an eye-in-hand camera. The code is available at https://github.com/HanjiangHu/camera-motion-smoothing.
Batch Normalization (BN) is a core and prevalent technique in accelerating the training of deep neural networks and improving the generalization on Computer Vision (CV) tasks. However, it fails to defend its position in Natural Language Processing (NLP), which is dominated by Layer Normalization (LN). In this paper, we are trying to answer why BN usually performs worse than LN in NLP tasks with Transformer models. We find that the inconsistency between training and inference of BN is the leading cause that results in the failure of BN in NLP. We define Training Inference Discrepancy (TID) to quantitatively measure this inconsistency and reveal that TID can indicate BN's performance, supported by extensive experiments, including image classification, neural machine translation, language modeling, sequence labeling, and text classification tasks. We find that BN can obtain much better test performance than LN when TID keeps small through training. To suppress the explosion of TID, we propose Regularized BN (RBN) that adds a simple regularization term to narrow the gap between batch statistics and population statistics of BN. RBN improves the performance of BN consistently and outperforms or is on par with LN on 17 out of 20 settings, involving ten datasets and two common variants of Transformer\footnote{Our code is available at \url{https://github.com/wjxts/RegularizedBN}}.
Data in real-world object detection often exhibits the long-tailed distribution. Existing solutions tackle this problem by mitigating the competition between the head and tail categories. However, due to the scarcity of training samples, tail categories are still unable to learn discriminative representations. Bringing more data into the training may alleviate the problem, but collecting instance-level annotations is an excruciating task. In contrast, image-level annotations are easily accessible but not fully exploited. In this paper, we propose a novel framework CLIS (multi-task Collaborative Learning with Image-level Supervision), which leverage image-level supervision to enhance the detection ability in a multi-task collaborative way. Specifically, there are an object detection task (consisting of an instance-classification task and a localization task) and an image-classification task in our framework, responsible for utilizing the two types of supervision. Different tasks are trained collaboratively by three key designs: (1) task-specialized sub-networks that learn specific representations of different tasks without feature entanglement. (2) a siamese sub-network for the image-classification task that shares its knowledge with the instance-classification task, resulting in feature enrichment of detectors. (3) a contrastive learning regularization that maintains representation consistency, bridging feature gaps of different supervision. Extensive experiments are conducted on the challenging LVIS dataset. Without sophisticated loss engineering, CLIS achieves an overall AP of 31.1 with 10.1 point improvement on tail categories, establishing a new state-of-the-art. Code will be at https://github.com/waveboo/CLIS.
Long-tail distribution is widely spread in real-world applications. Due to the extremely small ratio of instances, tail categories often show inferior accuracy. In this paper, we find such performance bottleneck is mainly caused by the imbalanced gradients, which can be categorized into two parts: (1) positive part, deriving from the samples of the same category, and (2) negative part, contributed by other categories. Based on comprehensive experiments, it is also observed that the gradient ratio of accumulated positives to negatives is a good indicator to measure how balanced a category is trained. Inspired by this, we come up with a gradient-driven training mechanism to tackle the long-tail problem: re-balancing the positive/negative gradients dynamically according to current accumulative gradients, with a unified goal of achieving balance gradient ratios. Taking advantage of the simple and flexible gradient mechanism, we introduce a new family of gradient-driven loss functions, namely equalization losses. We conduct extensive experiments on a wide spectrum of visual tasks, including two-stage/single-stage long-tailed object detection (LVIS), long-tailed image classification (ImageNet-LT, Places-LT, iNaturalist), and long-tailed semantic segmentation (ADE20K). Our method consistently outperforms the baseline models, demonstrating the effectiveness and generalization ability of the proposed equalization losses. Codes will be released at https://github.com/ModelTC/United-Perception.
The pretrain-finetune paradigm in modern computer vision facilitates the success of self-supervised learning, which tends to achieve better transferability than supervised learning. However, with the availability of massive labeled data, a natural question emerges: how to train a better model with both self and full supervision signals? In this paper, we propose Omni-suPErvised Representation leArning with hierarchical supervisions (OPERA) as a solution. We provide a unified perspective of supervisions from labeled and unlabeled data and propose a unified framework of fully supervised and self-supervised learning. We extract a set of hierarchical proxy representations for each image and impose self and full supervisions on the corresponding proxy representations. Extensive experiments on both convolutional neural networks and vision transformers demonstrate the superiority of OPERA in image classification, segmentation, and object detection. Code is available at: https://github.com/wangck20/OPERA.
Vision-language models such as CLIP are pretrained on large volumes of internet sourced image and text pairs, and have been shown to sometimes exhibit impressive zero- and low-shot image classification performance. However, due to their size, fine-tuning these models on new datasets can be prohibitively expensive, both in terms of the supervision and compute required. To combat this, a series of light-weight adaptation methods have been proposed to efficiently adapt such models when limited supervision is available. In this work, we show that while effective on internet-style datasets, even those remedies under-deliver on classification tasks with images that differ significantly from those commonly found online. To address this issue, we present a new approach called SVL-Adapter that combines the complementary strengths of both vision-language pretraining and self-supervised representation learning. We report an average classification accuracy improvement of 10% in the low-shot setting when compared to existing methods, on a set of challenging visual classification tasks. Further, we present a fully automatic way of selecting an important blending hyperparameter for our model that does not require any held-out labeled validation data. Code for our project is available here: https://github.com/omipan/svl_adapter.
When experience is scarce, models may have insufficient information to adapt to a new task. In this case, auxiliary information - such as a textual description of the task - can enable improved task inference and adaptation. In this work, we propose an extension to the Model-Agnostic Meta-Learning algorithm (MAML), which allows the model to adapt using auxiliary information as well as task experience. Our method, Fusion by Meta-Initialization (FuMI), conditions the model initialization on auxiliary information using a hypernetwork, rather than learning a single, task-agnostic initialization. Furthermore, motivated by the shortcomings of existing multi-modal few-shot learning benchmarks, we constructed iNat-Anim - a large-scale image classification dataset with succinct and visually pertinent textual class descriptions. On iNat-Anim, FuMI significantly outperforms uni-modal baselines such as MAML in the few-shot regime. The code for this project and a dataset exploration tool for iNat-Anim are publicly available at https://github.com/s-a-malik/multi-few .