Understanding whether self-supervised learning methods can scale with unlimited data is crucial for training large-scale models. In this work, we conduct an empirical study on the scaling capability of masked image modeling (MIM) methods (e.g., MAE) for visual recognition. Unlike most previous works that depend on the widely-used ImageNet dataset, which is manually curated and object-centric, we take a step further and propose to investigate this problem in a more practical setting. Specifically, we utilize the web-collected Coyo-700M dataset. We randomly sample varying numbers of training images from the Coyo dataset and construct a series of sub-datasets, containing 0.5M, 1M, 5M, 10M, and 100M images, for pre-training. Our goal is to investigate how the performance changes on downstream tasks when scaling with different sizes of data and models. The study reveals that: 1) MIM can be viewed as an effective method to improve the model capacity when the scale of the training data is relatively small; 2) Strong reconstruction targets can endow the models with increased capacities on downstream tasks; 3) MIM pre-training is data-agnostic under most scenarios, which means that the strategy of sampling pre-training data is non-critical. We hope these observations could provide valuable insights for future research on MIM.
相關內容
We present Emu, a Transformer-based multimodal foundation model, which can seamlessly generate images and texts in multimodal context. This omnivore model can take in any single-modality or multimodal data input indiscriminately (e.g., interleaved image, text and video) through a one-model-for-all autoregressive training process. First, visual signals are encoded into embeddings, and together with text tokens form an interleaved input sequence. Emu is then end-to-end trained with a unified objective of classifying the next text token or regressing the next visual embedding in the multimodal sequence. This versatile multimodality empowers the exploration of diverse pretraining data sources at scale, such as videos with interleaved frames and text, webpages with interleaved images and text, as well as web-scale image-text pairs and video-text pairs. Emu can serve as a generalist multimodal interface for both image-to-text and text-to-image tasks, and supports in-context image and text generation. Across a broad range of zero-shot/few-shot tasks including image captioning, visual question answering, video question answering and text-to-image generation, Emu demonstrates superb performance compared to state-of-the-art large multimodal models. Extended capabilities such as multimodal assistants via instruction tuning are also demonstrated with impressive performance.
Novelty detection in discrete sequences is a challenging task, since deviations from the process generating the normal data are often small or intentionally hidden. Novelties can be detected by modeling normal sequences and measuring the deviations of a new sequence from the model predictions. However, in many applications data is generated by several distinct processes so that models trained on all the data tend to over-generalize and novelties remain undetected. We propose to approach this challenge through decomposition: by clustering the data we break down the problem, obtaining simpler modeling task in each cluster which can be modeled more accurately. However, this comes at a trade-off, since the amount of training data per cluster is reduced. This is a particular problem for discrete sequences where state-of-the-art models are data-hungry. The success of this approach thus depends on the quality of the clustering, i.e., whether the individual learning problems are sufficiently simpler than the joint problem. While clustering discrete sequences automatically is a challenging and domain-specific task, it is often easy for human domain experts, given the right tools. In this paper, we adapt a state-of-the-art visual analytics tool for discrete sequence clustering to obtain informed clusters from domain experts and use LSTMs to model each cluster individually. Our extensive empirical evaluation indicates that this informed clustering outperforms automatic ones and that our approach outperforms state-of-the-art novelty detection methods for discrete sequences in three real-world application scenarios. In particular, decomposition outperforms a global model despite less training data on each individual cluster.
Neural radiance fields (NeRF) have shown great success in novel view synthesis. However, recovering high-quality details from real-world scenes is still challenging for the existing NeRF-based approaches, due to the potential imperfect calibration information and scene representation inaccuracy. Even with high-quality training frames, the synthetic novel views produced by NeRF models still suffer from notable rendering artifacts, such as noise and blur. To address this, we propose NeRFLiX, a general NeRF-agnostic restorer paradigm that learns a degradation-driven inter-viewpoint mixer. Specially, we design a NeRF-style degradation modeling approach and construct large-scale training data, enabling the possibility of effectively removing NeRF-native rendering artifacts for deep neural networks. Moreover, beyond the degradation removal, we propose an inter-viewpoint aggregation framework that fuses highly related high-quality training images, pushing the performance of cutting-edge NeRF models to entirely new levels and producing highly photo-realistic synthetic views. Based on this paradigm, we further present NeRFLiX++ with a stronger two-stage NeRF degradation simulator and a faster inter-viewpoint mixer, achieving superior performance with significantly improved computational efficiency. Notably, NeRFLiX++ is capable of restoring photo-realistic ultra-high-resolution outputs from noisy low-resolution NeRF-rendered views. Extensive experiments demonstrate the excellent restoration ability of NeRFLiX++ on various novel view synthesis benchmarks.
Despite their widespread adoption, neural conversation models have yet to exhibit natural chat capabilities with humans. In this research, we examine user utterances as causes and generated responses as effects, recognizing that changes in a cause should produce a different effect. To further explore this concept, we have compiled and expanded upon a new dataset called CausalDialogue through crowd-sourcing. This dataset includes multiple cause-effect pairs within a directed acyclic graph (DAG) structure. Our analysis reveals that traditional loss functions struggle to effectively incorporate the DAG structure, leading us to propose a causality-enhanced method called Exponential Maximum Average Treatment Effect (ExMATE) to enhance the impact of causality at the utterance level in training neural conversation models. To evaluate the needs of considering causality in dialogue generation, we built a comprehensive benchmark on CausalDialogue dataset using different models, inference, and training methods. Through experiments, we find that a causality-inspired loss like ExMATE can improve the diversity and agility of conventional loss function and there is still room for improvement to reach human-level quality on this new dataset.
Unsupervised/self-supervised representation learning in time series is critical since labeled samples are usually scarce in real-world scenarios. Existing approaches mainly leverage the contrastive learning framework, which automatically learns to understand the similar and dissimilar data pairs. Nevertheless, they are restricted to the prior knowledge of constructing pairs, cumbersome sampling policy, and unstable performances when encountering sampling bias. Also, few works have focused on effectively modeling across temporal-spectral relations to extend the capacity of representations. In this paper, we aim at learning representations for time series from a new perspective and propose Cross Reconstruction Transformer (CRT) to solve the aforementioned problems in a unified way. CRT achieves time series representation learning through a cross-domain dropping-reconstruction task. Specifically, we transform time series into the frequency domain and randomly drop certain parts in both time and frequency domains. Dropping can maximally preserve the global context compared to cropping and masking. Then a transformer architecture is utilized to adequately capture the cross-domain correlations between temporal and spectral information through reconstructing data in both domains, which is called Dropped Temporal-Spectral Modeling. To discriminate the representations in global latent space, we propose Instance Discrimination Constraint to reduce the mutual information between different time series and sharpen the decision boundaries. Additionally, we propose a specified curriculum learning strategy to optimize the CRT, which progressively increases the dropping ratio in the training process.
This work introduces a novel approach to pruning deep learning models by using distilled data. Unlike conventional strategies which primarily focus on architectural or algorithmic optimization, our method reconsiders the role of data in these scenarios. Distilled datasets capture essential patterns from larger datasets, and we demonstrate how to leverage this capability to enable a computationally efficient pruning process. Our approach can find sparse, trainable subnetworks (a.k.a. Lottery Tickets) up to 5x faster than Iterative Magnitude Pruning at comparable sparsity on CIFAR-10. The experimental results highlight the potential of using distilled data for resource-efficient neural network pruning, model compression, and neural architecture search.
The real-world data tends to be heavily imbalanced and severely skew the data-driven deep neural networks, which makes Long-Tailed Recognition (LTR) a massive challenging task. Existing LTR methods seldom train Vision Transformers (ViTs) with Long-Tailed (LT) data, while the off-the-shelf pretrain weight of ViTs always leads to unfair comparisons. In this paper, we systematically investigate the ViTs' performance in LTR and propose LiVT to train ViTs from scratch only with LT data. With the observation that ViTs suffer more severe LTR problems, we conduct Masked Generative Pretraining (MGP) to learn generalized features. With ample and solid evidence, we show that MGP is more robust than supervised manners. In addition, Binary Cross Entropy (BCE) loss, which shows conspicuous performance with ViTs, encounters predicaments in LTR. We further propose the balanced BCE to ameliorate it with strong theoretical groundings. Specially, we derive the unbiased extension of Sigmoid and compensate extra logit margins to deploy it. Our Bal-BCE contributes to the quick convergence of ViTs in just a few epochs. Extensive experiments demonstrate that with MGP and Bal-BCE, LiVT successfully trains ViTs well without any additional data and outperforms comparable state-of-the-art methods significantly, e.g., our ViT-B achieves 81.0% Top-1 accuracy in iNaturalist 2018 without bells and whistles. Code is available at //github.com/XuZhengzhuo/LiVT.
Deep models, e.g., CNNs and Vision Transformers, have achieved impressive achievements in many vision tasks in the closed world. However, novel classes emerge from time to time in our ever-changing world, requiring a learning system to acquire new knowledge continually. For example, a robot needs to understand new instructions, and an opinion monitoring system should analyze emerging topics every day. Class-Incremental Learning (CIL) enables the learner to incorporate the knowledge of new classes incrementally and build a universal classifier among all seen classes. Correspondingly, when directly training the model with new class instances, a fatal problem occurs -- the model tends to catastrophically forget the characteristics of former ones, and its performance drastically degrades. There have been numerous efforts to tackle catastrophic forgetting in the machine learning community. In this paper, we survey comprehensively recent advances in deep class-incremental learning and summarize these methods from three aspects, i.e., data-centric, model-centric, and algorithm-centric. We also provide a rigorous and unified evaluation of 16 methods in benchmark image classification tasks to find out the characteristics of different algorithms empirically. Furthermore, we notice that the current comparison protocol ignores the influence of memory budget in model storage, which may result in unfair comparison and biased results. Hence, we advocate fair comparison by aligning the memory budget in evaluation, as well as several memory-agnostic performance measures. The source code to reproduce these evaluations is available at //github.com/zhoudw-zdw/CIL_Survey/
Joint image-text embedding is the bedrock for most Vision-and-Language (V+L) tasks, where multimodality inputs are jointly processed for visual and textual understanding. In this paper, we introduce UNITER, a UNiversal Image-TExt Representation, learned through large-scale pre-training over four image-text datasets (COCO, Visual Genome, Conceptual Captions, and SBU Captions), which can power heterogeneous downstream V+L tasks with joint multimodal embeddings. We design three pre-training tasks: Masked Language Modeling (MLM), Image-Text Matching (ITM), and Masked Region Modeling (MRM, with three variants). Different from concurrent work on multimodal pre-training that apply joint random masking to both modalities, we use conditioned masking on pre-training tasks (i.e., masked language/region modeling is conditioned on full observation of image/text). Comprehensive analysis shows that conditioned masking yields better performance than unconditioned masking. We also conduct a thorough ablation study to find an optimal setting for the combination of pre-training tasks. Extensive experiments show that UNITER achieves new state of the art across six V+L tasks (over nine datasets), including Visual Question Answering, Image-Text Retrieval, Referring Expression Comprehension, Visual Commonsense Reasoning, Visual Entailment, and NLVR2.
It is a common paradigm in object detection frameworks to treat all samples equally and target at maximizing the performance on average. In this work, we revisit this paradigm through a careful study on how different samples contribute to the overall performance measured in terms of mAP. Our study suggests that the samples in each mini-batch are neither independent nor equally important, and therefore a better classifier on average does not necessarily mean higher mAP. Motivated by this study, we propose the notion of Prime Samples, those that play a key role in driving the detection performance. We further develop a simple yet effective sampling and learning strategy called PrIme Sample Attention (PISA) that directs the focus of the training process towards such samples. Our experiments demonstrate that it is often more effective to focus on prime samples than hard samples when training a detector. Particularly, On the MSCOCO dataset, PISA outperforms the random sampling baseline and hard mining schemes, e.g. OHEM and Focal Loss, consistently by more than 1% on both single-stage and two-stage detectors, with a strong backbone ResNeXt-101.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191