Large pre-trained multimodal models have demonstrated significant success in a range of downstream tasks, including image captioning, image-text retrieval, visual question answering (VQA), etc. However, many of these methods rely on image-text pairs collected from the web as pre-training data and unfortunately overlook the need for fine-grained feature alignment between vision and language modalities, which requires detailed understanding of images and language expressions. While integrating VQA and dense captioning (DC) into pre-training can address this issue, acquiring image-question-answer as well as image-location-caption triplets is challenging and time-consuming. Additionally, publicly available datasets for VQA and dense captioning are typically limited in scale due to manual data collection and labeling efforts. In this paper, we propose a novel method called Joint QA and DC GEneration (JADE), which utilizes a pre-trained multimodal model and easily-crawled image-text pairs to automatically generate and filter large-scale VQA and dense captioning datasets. We apply this method to the Conceptual Caption (CC3M) dataset to generate a new dataset called CC3M-QA-DC. Experiments show that when used for pre-training in a multi-task manner, CC3M-QA-DC can improve the performance with various backbones on various downstream tasks. Furthermore, our generated CC3M-QA-DC can be combined with larger image-text datasets (e.g., CC15M) and achieve competitive results compared with models using much more data. Code and dataset will be released.
Recently, model-based retrieval has emerged as a new paradigm in text retrieval that discards the index in the traditional retrieval model and instead memorizes the candidate corpora using model parameters. This design employs a sequence-to-sequence paradigm to generate document identifiers, which enables the complete capture of the relevance between queries and documents and simplifies the classic indexretrieval-rerank pipeline. Despite its attractive qualities, there remain several major challenges in model-based retrieval, including the discrepancy between pre-training and fine-tuning, and the discrepancy between training and inference. To deal with the above challenges, we propose a novel two-stage model-based retrieval approach called TOME, which makes two major technical contributions, including the utilization of tokenized URLs as identifiers and the design of a two-stage generation architecture. We also propose a number of training strategies to deal with the training difficulty as the corpus size increases. Extensive experiments and analysis on MS MARCO and Natural Questions demonstrate the effectiveness of our proposed approach, and we investigate the scaling laws of TOME by examining various influencing factors.
Diffusion models have recently dominated image synthesis and other related generative tasks. However, the iterative denoising process is expensive in computations at inference time, making diffusion models less practical for low-latency and scalable real-world applications. Post-training quantization of diffusion models can significantly reduce the model size and accelerate the sampling process without requiring any re-training. Nonetheless, applying existing post-training quantization methods directly to low-bit diffusion models can significantly impair the quality of generated samples. Specifically, for each denoising step, quantization noise leads to deviations in the estimated mean and mismatches with the predetermined variance schedule. Moreover, as the sampling process proceeds, the quantization noise may accumulate, resulting in a low signal-to-noise ratio (SNR) in late denoising steps. To address these challenges, we propose a unified formulation for the quantization noise and diffusion perturbed noise in the quantized denoising process. We first disentangle the quantization noise into its correlated and residual uncorrelated parts regarding its full-precision counterpart. The correlated part can be easily corrected by estimating the correlation coefficient. For the uncorrelated part, we calibrate the denoising variance schedule to absorb the excess variance resulting from quantization. Moreover, we propose a mixed-precision scheme to choose the optimal bitwidth for each denoising step, which prefers low bits to accelerate the early denoising steps while high bits maintain the high SNR for the late steps. Extensive experiments demonstrate that our method outperforms previous post-training quantized diffusion models in generating high-quality samples, with only a 0.06 increase in FID score compared to full-precision LDM-4 on ImageNet 256x256, while saving 19.9x bit operations.
Video anomaly detection (VAD) is a vital task with great practical applications in industrial surveillance, security system, and traffic control. Unlike previous unsupervised VAD methods that adopt a fixed structure to learn normality without considering different detection demands, we design a spatial-temporal hierarchical architecture (STHA) as a configurable architecture to flexibly detect different degrees of anomaly. The comprehensive structure of the STHA is delineated into a tripartite hierarchy, encompassing the following tiers: the stream level, the stack level, and the block level. Specifically, we design several auto-encoder-based blocks that possess varying capacities for extracting normal patterns. Then, we stack blocks according to the complexity degrees with both intra-stack and inter-stack residual links to learn hierarchical normality gradually. Considering the multisource knowledge of videos, we also model the spatial normality of video frames and temporal normality of RGB difference by designing two parallel streams consisting of stacks. Thus, STHA can provide various representation learning abilities by expanding or contracting hierarchically to detect anomalies of different degrees. Since the anomaly set is complicated and unbounded, our STHA can adjust its detection ability to adapt to the human detection demands and the complexity degree of anomaly that happened in the history of a scene. We conduct experiments on three benchmarks and perform extensive analysis, and the results demonstrate that our method performs comparablely to the state-of-the-art methods. In addition, we design a toy dataset to prove that our model can better balance the learning ability to adapt to different detection demands.
Attention-based contextual biasing approaches have shown significant improvements in the recognition of generic and/or personal rare-words in End-to-End Automatic Speech Recognition (E2E ASR) systems like neural transducers. These approaches employ cross-attention to bias the model towards specific contextual entities injected as bias-phrases to the model. Prior approaches typically relied on subword encoders for encoding the bias phrases. However, subword tokenizations are coarse and fail to capture granular pronunciation information which is crucial for biasing based on acoustic similarity. In this work, we propose to use lightweight character representations to encode fine-grained pronunciation features to improve contextual biasing guided by acoustic similarity between the audio and the contextual entities (termed acoustic biasing). We further integrate pretrained neural language model (NLM) based encoders to encode the utterance's semantic context along with contextual entities to perform biasing informed by the utterance's semantic context (termed semantic biasing). Experiments using a Conformer Transducer model on the Librispeech dataset show a 4.62% - 9.26% relative WER improvement on different biasing list sizes over the baseline contextual model when incorporating our proposed acoustic and semantic biasing approach. On a large-scale in-house dataset, we observe 7.91% relative WER improvement compared to our baseline model. On tail utterances, the improvements are even more pronounced with 36.80% and 23.40% relative WER improvements on Librispeech rare words and an in-house testset respectively.
Industrial image anomaly detection under the setting of one-class classification has significant practical value. However, most existing models struggle to extract separable feature representations when performing feature embedding and struggle to build compact descriptions of normal features when performing one-class classification. One direct consequence of this is that most models perform poorly in detecting logical anomalies which violate contextual relationships. Focusing on more effective and comprehensive anomaly detection, we propose a network based on self-supervised learning and self-attentive graph convolution (SLSG) for anomaly detection. SLSG uses a generative pre-training network to assist the encoder in learning the embedding of normal patterns and the reasoning of position relationships. Subsequently, SLSG introduces the pseudo-prior knowledge of anomaly through simulated abnormal samples. By comparing the simulated anomalies, SLSG can better summarize the normal features and narrow down the hypersphere used for one-class classification. In addition, with the construction of a more general graph structure, SLSG comprehensively models the dense and sparse relationships among elements in the image, which further strengthens the detection of logical anomalies. Extensive experiments on benchmark datasets show that SLSG achieves superior anomaly detection performance, demonstrating the effectiveness of our method.
Learning to optimize (L2O) has emerged as a powerful framework for black-box optimization (BBO). L2O learns the optimization strategies from the target task automatically without human intervention. This paper focuses on obtaining better performance when handling high-dimensional and expensive BBO with little function evaluation cost, which is the core challenge of black-box optimization. However, current L2O-based methods are weak for this due to a large number of evaluations on expensive black-box functions during training and poor representation of optimization strategy. To achieve this, 1) we utilize the cheap surrogate functions of the target task to guide the design of the optimization strategies; 2) drawing on the mechanism of evolutionary algorithm (EA), we propose a novel framework called B2Opt, which has a stronger representation of optimization strategies. Compared to the BBO baselines, B2Opt can achieve 3 to $10^6$ times performance improvement with less function evaluation cost. We test our proposal in high-dimensional synthetic functions and two real-world applications. We also find that deep B2Opt performs better than shallow ones.
Evolutionary algorithms (EAs) have emerged as a powerful framework for expensive black-box optimization. Obtaining better solutions with less computational cost is essential and challenging for black-box optimization. The most critical obstacle is figuring out how to effectively use the target task information to form an efficient optimization strategy. However, current methods are weak due to the poor representation of the optimization strategy and the inefficient interaction between the optimization strategy and the target task. To overcome the above limitations, we design a learned EA (LEA) to realize the move from hand-designed optimization strategies to learned optimization strategies, including not only hyperparameters but also update rules. Unlike traditional EAs, LEA has high adaptability to the target task and can obtain better solutions with less computational cost. LEA is also able to effectively utilize the low-fidelity information of the target task to form an efficient optimization strategy. The experimental results on one synthetic case, CEC 2013, and two real-world cases show the advantages of learned optimization strategies over human-designed baselines. In addition, LEA is friendly to the acceleration provided by Graphics Processing Units and runs 102 times faster than unaccelerated EA when evolving 32 populations, each containing 6400 individuals.
In this paper, we propose a Vision-Audio-Language Omni-peRception pretraining model (VALOR) for multi-modal understanding and generation. Different from widely-studied vision-language pretraining models, VALOR jointly models relationships of vision, audio and language in an end-to-end manner. It contains three separate encoders for single modality representations, and a decoder for multimodal conditional text generation. We design two pretext tasks to pretrain VALOR model, including Multimodal Grouping Alignment (MGA) and Multimodal Grouping Captioning (MGC). MGA projects vision, language and audio to the same common space, building vision-language, audio-language and audiovisual-language alignment simultaneously. MGC learns how to generate text tokens in conditions of vision, audio or their both. To promote vision-audio-language pretraining research, we construct a large-scale high-quality tri-modality dataset named VALOR-1M, which contains 1M audiable videos with human annotated audiovisual captions. Extensive experiments show that VALOR can learn strong multimodal correlations and be generalized to various downstream tasks (e.g., retrieval, captioning and question answering), with different input modalities (e.g., vision-language, audio-language and audiovisual-language). VALOR achieves new state-of-the-art performances on series of public cross-modality benchmarks. Code and data are available at project page https://casia-iva-group.github.io/projects/VALOR.
Video anomaly detection (VAD) is a significant computer vision problem. Existing deep neural network (DNN) based VAD methods mostly follow the route of frame reconstruction or frame prediction. However, the lack of mining and learning of higher-level visual features and temporal context relationships in videos limits the further performance of these two approaches. Inspired by video codec theory, we introduce a brand-new VAD paradigm to break through these limitations: First, we propose a new task of video event restoration based on keyframes. Encouraging DNN to infer missing multiple frames based on video keyframes so as to restore a video event, which can more effectively motivate DNN to mine and learn potential higher-level visual features and comprehensive temporal context relationships in the video. To this end, we propose a novel U-shaped Swin Transformer Network with Dual Skip Connections (USTN-DSC) for video event restoration, where a cross-attention and a temporal upsampling residual skip connection are introduced to further assist in restoring complex static and dynamic motion object features in the video. In addition, we propose a simple and effective adjacent frame difference loss to constrain the motion consistency of the video sequence. Extensive experiments on benchmarks demonstrate that USTN-DSC outperforms most existing methods, validating the effectiveness of our method.