Table structure recognition (TSR) aims to convert tabular images into a machine-readable format, where a visual encoder extracts image features and a textual decoder generates table-representing tokens. Existing approaches use classic convolutional neural network (CNN) backbones for the visual encoder and transformers for the textual decoder. However, this hybrid CNN-Transformer architecture introduces a complex visual encoder that accounts for nearly half of the total model parameters, markedly reduces both training and inference speed, and hinders the potential for self-supervised learning in TSR. In this work, we design a lightweight visual encoder for TSR without sacrificing expressive power. We discover that a convolutional stem can match classic CNN backbone performance, with a much simpler model. The convolutional stem strikes an optimal balance between two crucial factors for high-performance TSR: a higher receptive field (RF) ratio and a longer sequence length. This allows it to "see" an appropriate portion of the table and "store" the complex table structure within sufficient context length for the subsequent transformer. We conducted reproducible ablation studies and open-sourced our code at https://github.com/poloclub/tsr-convstem to enhance transparency, inspire innovations, and facilitate fair comparisons in our domain as tables are a promising modality for representation learning.
Minimum Bayes Risk (MBR) decoding has been shown to be a powerful alternative to beam search decoding in a variety of text generation tasks. MBR decoding selects a hypothesis from a pool of hypotheses that has the least expected risk under a probability model according to a given utility function. Since it is impractical to compute the expected risk exactly over all possible hypotheses, two approximations are commonly used in MBR. First, it integrates over a sampled set of hypotheses rather than over all possible hypotheses. Second, it estimates the probability of each hypothesis using a Monte Carlo estimator. While the first approximation is necessary to make it computationally feasible, the second is not essential since we typically have access to the model probability at inference time. We propose Model-Based MBR (MBMBR), a variant of MBR that uses the model probability itself as the estimate of the probability distribution instead of the Monte Carlo estimate. We show analytically and empirically that the model-based estimate is more promising than the Monte Carlo estimate in text generation tasks. Our experiments show that MBMBR outperforms MBR in several text generation tasks, both with encoder-decoder models and with large language models.
Recently, unified speech-text models, such as SpeechGPT, VioLA, and AudioPaLM, have achieved remarkable performance on various speech tasks. These models discretize speech signals into tokens (speech discretization) and use a shared vocabulary for both text and speech tokens. Then they train a single decoder-only Transformer on a mixture of speech tasks. However, these models rely on the Loss Masking strategy for the ASR task, which ignores the dependency among speech tokens. In this paper, we propose to model speech tokens in an autoregressive way, similar to text. We find that applying the conventional cross-entropy loss on input speech tokens does not consistently improve the ASR performance over the Loss Masking approach. To address this issue, we propose a novel approach denoted Smoothed Label Distillation (SLD), which applies a KL divergence loss with smoothed labels on speech tokens. Our experiments show that SLD effectively models speech tokens and outperforms Loss Masking for decoder-only Transformers in ASR tasks with different speech discretization methods. The source code can be found here: https://github.com/alibaba-damo-academy/SpokenNLP/tree/main/sld
Developing a new Salient Object Detection (SOD) model involves selecting an ImageNet pre-trained backbone and creating novel feature refinement modules to use backbone features. However, adding new components to a pre-trained backbone needs retraining the whole network on the ImageNet dataset, which requires significant time. Hence, we explore developing a neural network from scratch directly trained on SOD without ImageNet pre-training. Such a formulation offers full autonomy to design task-specific components. To that end, we propose SODAWideNet, an encoder-decoder-style network for Salient Object Detection. We deviate from the commonly practiced paradigm of narrow and deep convolutional models to a wide and shallow architecture, resulting in a parameter-efficient deep neural network. To achieve a shallower network, we increase the receptive field from the beginning of the network using a combination of dilated convolutions and self-attention. Therefore, we propose Multi Receptive Field Feature Aggregation Module (MRFFAM) that efficiently obtains discriminative features from farther regions at higher resolutions using dilated convolutions. Next, we propose Multi-Scale Attention (MSA), which creates a feature pyramid and efficiently computes attention across multiple resolutions to extract global features from larger feature maps. Finally, we propose two variants, SODAWideNet-S (3.03M) and SODAWideNet (9.03M), that achieve competitive performance against state-of-the-art models on five datasets.
Multi-modal Large Language Models (MLLMs) have demonstrated impressive instruction abilities across various open-ended tasks. However, previous methods primarily focus on enhancing multi-modal capabilities. In this work, we introduce a versatile multi-modal large language model, mPLUG-Owl2, which effectively leverages modality collaboration to improve performance in both text and multi-modal tasks. mPLUG-Owl2 utilizes a modularized network design, with the language decoder acting as a universal interface for managing different modalities. Specifically, mPLUG-Owl2 incorporates shared functional modules to facilitate modality collaboration and introduces a modality-adaptive module that preserves modality-specific features. Extensive experiments reveal that mPLUG-Owl2 is capable of generalizing both text tasks and multi-modal tasks and achieving state-of-the-art performances with a single generic model. Notably, mPLUG-Owl2 is the first MLLM model that demonstrates the modality collaboration phenomenon in both pure-text and multi-modal scenarios, setting a pioneering path in the development of future multi-modal foundation models.
Deepfake videos, generated through AI faceswapping techniques, have garnered considerable attention due to their potential for powerful impersonation attacks. While existing research primarily focuses on binary classification to discern between real and fake videos, however determining the specific generation model for a fake video is crucial for forensic investigation. Addressing this gap, this paper investigates the model attribution problem of Deepfake videos from a recently proposed dataset, Deepfakes from Different Models (DFDM), derived from various Autoencoder models. The dataset comprises 6,450 Deepfake videos generated by five distinct models with variations in encoder, decoder, intermediate layer, input resolution, and compression ratio. This study formulates Deepfakes model attribution as a multiclass classification task, proposing a segment of VGG19 as a feature extraction backbone, known for its effectiveness in imagerelated tasks, while integrated a Capsule Network with a Spatio-Temporal attention mechanism. The Capsule module captures intricate hierarchies among features for robust identification of deepfake attributes. Additionally, the video-level fusion technique leverages temporal attention mechanisms to handle concatenated feature vectors, capitalizing on inherent temporal dependencies in deepfake videos. By aggregating insights across frames, our model gains a comprehensive understanding of video content, resulting in more precise predictions. Experimental results on the deepfake benchmark dataset (DFDM) demonstrate the efficacy of our proposed method, achieving up to a 4% improvement in accurately categorizing deepfake videos compared to baseline models while demanding fewer computational resources.
We present a novel usage of Transformers to make image classification interpretable. Unlike mainstream classifiers that wait until the last fully connected layer to incorporate class information to make predictions, we investigate a proactive approach, asking each class to search for itself in an image. We realize this idea via a Transformer encoder-decoder inspired by DEtection TRansformer (DETR). We learn "class-specific" queries (one for each class) as input to the decoder, enabling each class to localize its patterns in an image via cross-attention. We name our approach INterpretable TRansformer (INTR), which is fairly easy to implement and exhibits several compelling properties. We show that INTR intrinsically encourages each class to attend distinctively; the cross-attention weights thus provide a faithful interpretation of the prediction. Interestingly, via "multi-head" cross-attention, INTR could identify different "attributes" of a class, making it particularly suitable for fine-grained classification and analysis, which we demonstrate on eight datasets. Our code and pre-trained models are publicly accessible at the Imageomics Institute GitHub site: https://github.com/Imageomics/INTR.
Contemporary translation engines based on the encoder-decoder framework have made significant strides in development. However, the emergence of Large Language Models (LLMs) has disrupted their position by presenting the potential for achieving superior translation quality. To uncover the circumstances in which LLMs excel and explore how their strengths can be harnessed to enhance translation quality, we first conduct a comprehensive analysis to assess the strengths and limitations of various commercial NMT systems and MT-oriented LLMs. Our findings indicate that neither NMT nor MT-oriented LLMs alone can effectively address all the translation issues, but MT-oriented LLMs show promise as a complementary solution to NMT systems. Building upon these insights, we propose Cooperative Decoding (CoDec), which treats NMT systems as a pretranslation model and MT-oriented LLMs as a supplemental solution to handle complex scenarios beyond the capability of NMT alone. Experimental results on the WMT22 test sets and a newly collected test set WebCrawl demonstrate the effectiveness and efficiency of CoDec, highlighting its potential as a robust solution for combining NMT systems with MT-oriented LLMs in the field of machine translation.
In this paper, we unveil that Language Models (LMs) can acquire new capabilities by assimilating parameters from homologous models without retraining or GPUs. We first introduce DARE to set most delta parameters (i.e., the disparity between fine-tuned and pre-trained parameters) to zeros without affecting the abilities of Supervised Fine-Tuning (SFT) LMs, which randomly Drops delta parameters with a ratio $p$ And REscales the remaining ones by $1 / (1 - p)$ to approximate the original embeddings. Then, we use DARE as a versatile plug-in to sparsify delta parameters of multiple SFT homologous models for mitigating parameter interference and merge them into a single model by parameter fusing. We experiment with encoder- and decoder-based LMs, showing that: (1) SFT delta parameter value ranges are typically small (within 0.002) with extreme redundancy, and DARE can effortlessly eliminate 90% or even 99% of them; (2) DARE can merge multiple task-specific LMs into one LM with diverse capabilities. Notably, this phenomenon is more pronounced in large-scale LMs, where the merged LM reveals the potential to surpass the performance of any source LM, providing a new discovery. We also utilize DARE to create a merged LM that ranks first among models with 7 billion parameters on the Open LLM Leaderboard.
This paper presents a novel supervised convolutional neural network architecture, "DUCK-Net", capable of effectively learning and generalizing from small amounts of medical images to perform accurate segmentation tasks. Our model utilizes an encoder-decoder structure with a residual downsampling mechanism and a custom convolutional block to capture and process image information at multiple resolutions in the encoder segment. We employ data augmentation techniques to enrich the training set, thus increasing our model's performance. While our architecture is versatile and applicable to various segmentation tasks, in this study, we demonstrate its capabilities specifically for polyp segmentation in colonoscopy images. We evaluate the performance of our method on several popular benchmark datasets for polyp segmentation, Kvasir-SEG, CVC-ClinicDB, CVC-ColonDB, and ETIS-LARIBPOLYPDB showing that it achieves state-of-the-art results in terms of mean Dice coefficient, Jaccard index, Precision, Recall, and Accuracy. Our approach demonstrates strong generalization capabilities, achieving excellent performance even with limited training data. The code is publicly available on GitHub: https://github.com/RazvanDu/DUCK-Net
In this paper, we describe the University of Lyon 2 submission to the Strict-Small track of the BabyLM competition. The shared task is created with an emphasis on small-scale language modelling from scratch on limited-size data and human language acquisition. Dataset released for the Strict-Small track has 10M words, which is comparable to children's vocabulary size. We approach the task with an architecture search, minimizing masked language modelling loss on the data of the shared task. Having found an optimal configuration, we introduce two small-size language models (LMs) that were submitted for evaluation, a 4-layer encoder with 8 attention heads and a 6-layer decoder model with 12 heads which we term Bebeshka and Zlata, respectively. Despite being half the scale of the baseline LMs, our proposed models achieve comparable performance. We further explore the applicability of small-scale language models in tasks involving moral judgment, aligning their predictions with human values. These findings highlight the potential of compact LMs in addressing practical language understanding tasks.
A simple design recipe for deep Transformers is to compose identical building blocks. But standard transformer blocks are far from simple, interweaving attention and MLP sub-blocks with skip connections & normalisation layers in precise arrangements. This complexity leads to brittle architectures, where seemingly minor changes can significantly reduce training speed, or render models untrainable. In this work, we ask to what extent the standard transformer block can be simplified? Combining signal propagation theory and empirical observations, we motivate modifications that allow many block components to be removed with no loss of training speed, including skip connections, projection or value parameters, sequential sub-blocks and normalisation layers. In experiments on both autoregressive decoder-only and BERT encoder-only models, our simplified transformers emulate the per-update training speed and performance of standard transformers, while enjoying 15% faster training throughput, and using 15% fewer parameters.
Video anomaly detection deals with the recognition of abnormal events in videos. Apart from the visual signal, video anomaly detection has also been addressed with the use of skeleton sequences. We propose a holistic representation of skeleton trajectories to learn expected motions across segments at different times. Our approach uses multitask learning to reconstruct any continuous unobserved temporal segment of the trajectory allowing the extrapolation of past or future segments and the interpolation of in-between segments. We use an end-to-end attention-based encoder-decoder. We encode temporally occluded trajectories, jointly learn latent representations of the occluded segments, and reconstruct trajectories based on expected motions across different temporal segments. Extensive experiments on three trajectory-based video anomaly detection datasets show the advantages and effectiveness of our approach with state-of-the-art results on anomaly detection in skeleton trajectories.
This paper presents an approach for compressing point cloud geometry by leveraging a lightweight super-resolution network. The proposed method involves decomposing a point cloud into a base point cloud and the interpolation patterns for reconstructing the original point cloud. While the base point cloud can be efficiently compressed using any lossless codec, such as Geometry-based Point Cloud Compression, a distinct strategy is employed for handling the interpolation patterns. Rather than directly compressing the interpolation patterns, a lightweight super-resolution network is utilized to learn this information through overfitting. Subsequently, the network parameter is transmitted to assist in point cloud reconstruction at the decoder side. Notably, our approach differentiates itself from lookup table-based methods, allowing us to obtain more accurate interpolation patterns by accessing a broader range of neighboring voxels at an acceptable computational cost. Experiments on MPEG Cat1 (Solid) and Cat2 datasets demonstrate the remarkable compression performance achieved by our method.
Learning representations from videos requires understanding continuous motion and visual correspondences between frames. In this paper, we introduce the Concatenated Masked Autoencoders (CatMAE) as a spatial-temporal learner for self-supervised video representation learning. For the input sequence of video frames, CatMAE keeps the initial frame unchanged while applying substantial masking (95%) to subsequent frames. The encoder in CatMAE is responsible for encoding visible patches for each frame individually; subsequently, for each masked frame, the decoder leverages visible patches from both previous and current frames to reconstruct the original image. Our proposed method enables the model to estimate the motion information between visible patches, match the correspondences between preceding and succeeding frames, and ultimately learn the evolution of scenes. Furthermore, we propose a new data augmentation strategy, Video-Reverse (ViRe), which uses reversed video frames as the model's reconstruction targets. This further encourages the model to utilize continuous motion details and correspondences to complete the reconstruction, thereby enhancing the model's capabilities. Compared to the most advanced pre-training methods, CatMAE achieves a leading level in video segmentation tasks and action recognition tasks.
This paper presents FlowSUM, a normalizing flows-based variational encoder-decoder framework for Transformer-based summarization. Our approach tackles two primary challenges in variational summarization: insufficient semantic information in latent representations and posterior collapse during training. To address these challenges, we employ normalizing flows to enable flexible latent posterior modeling, and we propose a controlled alternate aggressive training (CAAT) strategy with an improved gate mechanism. Experimental results show that FlowSUM significantly enhances the quality of generated summaries and unleashes the potential for knowledge distillation with minimal impact on inference time. Furthermore, we investigate the issue of posterior collapse in normalizing flows and analyze how the summary quality is affected by the training strategy, gate initialization, and the type and number of normalizing flows used, offering valuable insights for future research.
One of the common and promising deep learning approaches used for medical image segmentation is transformers, as they can capture long-range dependencies among the pixels by utilizing self-attention. Despite being successful in medical image segmentation, transformers face limitations in capturing local contexts of pixels in multimodal dimensions. We propose a Medical Image Segmentation Transformer (MIST) incorporating a novel Convolutional Attention Mixing (CAM) decoder to address this issue. MIST has two parts: a pre-trained multi-axis vision transformer (MaxViT) is used as an encoder, and the encoded feature representation is passed through the CAM decoder for segmenting the images. In the CAM decoder, an attention-mixer combining multi-head self-attention, spatial attention, and squeeze and excitation attention modules is introduced to capture long-range dependencies in all spatial dimensions. Moreover, to enhance spatial information gain, deep and shallow convolutions are used for feature extraction and receptive field expansion, respectively. The integration of low-level and high-level features from different network stages is enabled by skip connections, allowing MIST to suppress unnecessary information. The experiments show that our MIST transformer with CAM decoder outperforms the state-of-the-art models specifically designed for medical image segmentation on the ACDC and Synapse datasets. Our results also demonstrate that adding the CAM decoder with a hierarchical transformer improves segmentation performance significantly. Our model with data and code is publicly available on GitHub.
In medical image segmentation, domain generalization poses a significant challenge due to domain shifts caused by variations in data acquisition devices and other factors. These shifts are particularly pronounced in the most common scenario, which involves only single-source domain data due to privacy concerns. To address this, we draw inspiration from the self-supervised learning paradigm that effectively discourages overfitting to the source domain. We propose the Denoising Y-Net (DeY-Net), a novel approach incorporating an auxiliary denoising decoder into the basic U-Net architecture. The auxiliary decoder aims to perform denoising training, augmenting the domain-invariant representation that facilitates domain generalization. Furthermore, this paradigm provides the potential to utilize unlabeled data. Building upon denoising training, we propose Denoising Test Time Adaptation (DeTTA) that further: (i) adapts the model to the target domain in a sample-wise manner, and (ii) adapts to the noise-corrupted input. Extensive experiments conducted on widely-adopted liver segmentation benchmarks demonstrate significant domain generalization improvements over our baseline and state-of-the-art results compared to other methods. Code is available at https://github.com/WenRuxue/DeTTA.
Multi-horizon time series forecasting, crucial across diverse domains, demands high accuracy and speed. While AutoRegressive (AR) models excel in short-term predictions, they suffer speed and error issues as the horizon extends. Non-AutoRegressive (NAR) models suit long-term predictions but struggle with interdependence, yielding unrealistic results. We introduce AMLNet, an innovative NAR model that achieves realistic forecasts through an online Knowledge Distillation (KD) approach. AMLNet harnesses the strengths of both AR and NAR models by training a deep AR decoder and a deep NAR decoder in a collaborative manner, serving as ensemble teachers that impart knowledge to a shallower NAR decoder. This knowledge transfer is facilitated through two key mechanisms: 1) outcome-driven KD, which dynamically weights the contribution of KD losses from the teacher models, enabling the shallow NAR decoder to incorporate the ensemble's diversity; and 2) hint-driven KD, which employs adversarial training to extract valuable insights from the model's hidden states for distillation. Extensive experimentation showcases AMLNet's superiority over conventional AR and NAR models, thereby presenting a promising avenue for multi-horizon time series forecasting that enhances accuracy and expedites computation.
Visual object tracking is a fundamental component of transportation systems, especially for intelligent driving. Despite achieving state-of-the-art performance in visual tracking, recent single-branch trackers tend to overlook the weak prior assumptions associated with the Vision Transformer (ViT) encoder and inference pipeline in visual tracking. Moreover, the effectiveness of discriminative trackers remains constrained due to the adoption of the dual-branch pipeline. To tackle the inferior effectiveness of vanilla ViT, we propose an Adaptive ViT Model Prediction tracker (AViTMP) to design a customised tracking method. This method bridges the single-branch network with discriminative models for the first time. Specifically, in the proposed encoder AViT encoder, we introduce a tracking-tailored Adaptor module for vanilla ViT and a joint target state embedding to enrich the target-prior embedding paradigm. Then, we combine the AViT encoder with a discriminative transformer-specific model predictor to predict the accurate location. Furthermore, to mitigate the limitations of conventional inference practice, we present a novel inference pipeline called CycleTrack, which bolsters the tracking robustness in the presence of distractors via bidirectional cycle tracking verification. In the experiments, we evaluated AViTMP on eight tracking benchmarks for a comprehensive assessment, including LaSOT, LaSOTExtSub, AVisT, etc. The experimental results unequivocally establish that, under fair comparison, AViTMP achieves state-of-the-art performance, especially in terms of long-term tracking and robustness. The source code will be released at https://github.com/Tchuanm/AViTMP.