U-Nets are a go-to, state-of-the-art neural architecture across numerous tasks for continuous signals on a square such as images and Partial Differential Equations (PDE), however their design and architecture is understudied. In this paper, we provide a framework for designing and analysing general U-Net architectures. We present theoretical results which characterise the role of the encoder and decoder in a U-Net, their high-resolution scaling limits and their conjugacy to ResNets via preconditioning. We propose Multi-ResNets, U-Nets with a simplified, wavelet-based encoder without learnable parameters. Further, we show how to design novel U-Net architectures which encode function constraints, natural bases, or the geometry of the data. In diffusion models, our framework enables us to identify that high-frequency information is dominated by noise exponentially faster, and show how U-Nets with average pooling exploit this. In our experiments, we demonstrate how Multi-ResNets achieve competitive and often superior performance compared to classical U-Nets in image segmentation, PDE surrogate modelling, and generative modelling with diffusion models. Our U-Net framework paves the way to study the theoretical properties of U-Nets and design natural, scalable neural architectures for a multitude of problems beyond the square.
One of the most important challenges in text generation systems is to produce outputs that are not only correct but also diverse. Recently, Minimum Bayes-Risk (MBR) decoding has gained prominence for generating sentences of the highest quality among the decoding algorithms. However, existing algorithms proposed for generating diverse outputs are predominantly based on beam search or random sampling, thus their output quality is capped by these underlying methods. In this paper, we investigate an alternative approach -- we develop diversity-promoting decoding algorithms by enforcing diversity objectives to MBR decoding. We propose two variants of MBR, Diverse MBR (DMBR) and $k$-medoids MBR (KMBR), methods to generate a set of sentences with high quality and diversity. We evaluate DMBR and KMBR on a variety of directed text generation tasks using encoder-decoder models and a large language model with prompting. The experimental results show that the proposed method achieves a better trade-off than the diverse beam search and sampling algorithms.
The problem of audio-to-audio (A2A) style transfer involves replacing the style features of the source audio with those from the target audio while preserving the content related attributes of the source audio. In this paper, we propose an efficient approach, termed as Zero-shot Emotion Style Transfer (ZEST), that allows the transfer of emotional content present in the given source audio with the one embedded in the target audio while retaining the speaker and speech content from the source. The proposed system builds upon decomposing speech into semantic tokens, speaker representations and emotion embeddings. Using these factors, we propose a framework to reconstruct the pitch contour of the given speech signal and train a decoder that reconstructs the speech signal. The model is trained using a self-supervision based reconstruction loss. During conversion, the emotion embedding is alone derived from the target audio, while rest of the factors are derived from the source audio. In our experiments, we show that, even without using parallel training data or labels from the source or target audio, we illustrate zero shot emotion transfer capabilities of the proposed ZEST model using objective and subjective quality evaluations.
Omnidirectional stereo matching (OSM) is an essential and reliable means for $360^{\circ}$ depth sensing. However, following earlier works on conventional stereo matching, prior state-of-the-art (SOTA) methods rely on a 3D encoder-decoder block to regularize the cost volume, causing the whole system complicated and sub-optimal results. Recently, the Recurrent All-pairs Field Transforms (RAFT) based approach employs the recurrent update in 2D and has efficiently improved image-matching tasks, ie, optical flow, and stereo matching. To bridge the gap between OSM and RAFT, we mainly propose an opposite adaptive weighting scheme to seamlessly transform the outputs of spherical sweeping of OSM into the required inputs for the recurrent update, thus creating a recurrent omnidirectional stereo matching (RomniStereo) algorithm. Furthermore, we introduce two techniques, ie, grid embedding and adaptive context feature generation, which also contribute to RomniStereo's performance. Our best model improves the average MAE metric by 40.7\% over the previous SOTA baseline across five datasets. When visualizing the results, our models demonstrate clear advantages on both synthetic and realistic examples. The code is available at \url{https://github.com/HalleyJiang/RomniStereo}.
End-to-end multi-talker speech recognition has garnered great interest as an effective approach to directly transcribe overlapped speech from multiple speakers. Current methods typically adopt either 1) single-input multiple-output (SIMO) models with a branched encoder, or 2) single-input single-output (SISO) models based on attention-based encoder-decoder architecture with serialized output training (SOT). In this work, we propose a Cross-Speaker Encoding (CSE) network to address the limitations of SIMO models by aggregating cross-speaker representations. Furthermore, the CSE model is integrated with SOT to leverage both the advantages of SIMO and SISO while mitigating their drawbacks. To the best of our knowledge, this work represents an early effort to integrate SIMO and SISO for multi-talker speech recognition. Experiments on the two-speaker LibrispeechMix dataset show that the CES model reduces word error rate (WER) by 8% over the SIMO baseline. The CSE-SOT model reduces WER by 10% overall and by 16% on high-overlap speech compared to the SOT model. Code is available at https://github.com/kjw11/CSEnet-ASR.
Diffusion models are the current state of the art for generating photorealistic images. Controlling the sampling process for constrained image generation tasks such as inpainting, however, remains challenging since exact conditioning on such constraints is intractable. While existing methods use various techniques to approximate the constrained posterior, this paper proposes to exploit the ability of Tractable Probabilistic Models (TPMs) to exactly and efficiently compute the constrained posterior, and to leverage this signal to steer the denoising process of diffusion models. Specifically, this paper adopts a class of expressive TPMs termed Probabilistic Circuits (PCs). Building upon prior advances, we further scale up PCs and make them capable of guiding the image generation process of diffusion models. Empirical results suggest that our approach can consistently improve the overall quality and semantic coherence of inpainted images across three natural image datasets (i.e., CelebA-HQ, ImageNet, and LSUN) with only $\sim\! 10 \%$ additional computational overhead brought by the TPM. Further, with the help of an image encoder and decoder, our method can readily accept semantic constraints on specific regions of the image, which opens up the potential for more controlled image generation tasks. In addition to proposing a new framework for constrained image generation, this paper highlights the benefit of more tractable models and motivates the development of expressive TPMs.
Low-density parity-check (LDPC) codes have been successfully commercialized in communication systems due to their strong error correction capabilities and simple decoding process. However, the error-floor phenomenon of LDPC codes, in which the error rate stops decreasing rapidly at a certain level, presents challenges for achieving extremely low error rates and deploying LDPC codes in scenarios demanding ultra-high reliability. In this work, we propose training methods for neural min-sum (NMS) decoders to eliminate the error-floor effect. First, by leveraging the boosting learning technique of ensemble networks, we divide the decoding network into two neural decoders and train the post decoder to be specialized for uncorrected words that the first decoder fails to correct. Secondly, to address the vanishing gradient issue in training, we introduce a block-wise training schedule that locally trains a block of weights while retraining the preceding block. Lastly, we show that assigning different weights to unsatisfied check nodes effectively lowers the error-floor with a minimal number of weights. By applying these training methods to standard LDPC codes, we achieve the best error-floor performance compared to other decoding methods. The proposed NMS decoder, optimized solely through novel training methods without additional modules, can be integrated into existing LDPC decoders without incurring extra hardware costs. The source code is available at https://github.com/ghy1228/LDPC_Error_Floor .
Large language models (LLMs) have made significant advancements in code-related tasks, yet many LLMs treat code as simple sequences, neglecting its structured nature. We introduce AST-T5, a novel pretraining paradigm that leverages the Abstract Syntax Tree (AST) for enhanced code generation, transpilation, and understanding. Using dynamic programming, our AST-Aware Segmentation retains code structure, while our AST-Aware Span Corruption objective equips the model to reconstruct various code structures. Unlike other models, AST-T5 avoids intricate program analyses or architectural changes, so it integrates seamlessly with any encoder-decoder Transformer. Evaluations show that AST-T5 consistently outperforms similar-sized LMs across various code-related tasks. Structure-awareness makes AST-T5 particularly powerful in code-to-code tasks, surpassing CodeT5 by 2 points in exact match score for the Bugs2Fix task and by 3 points in exact match score for Java-C# Transpilation in CodeXGLUE. Our code and model are publicly available at https://github.com/gonglinyuan/ast_t5.
We present WidthFormer, a novel transformer-based module to compute Bird's-Eye-View (BEV) representations from multi-view cameras for real-time autonomous-driving applications. WidthFormer is computationally efficient, robust and does not require any special engineering effort to deploy. We first introduce a novel 3D positional encoding mechanism capable of accurately encapsulating 3D geometric information, which enables our model to compute high-quality BEV representations with only a single transformer decoder layer. This mechanism is also beneficial for existing sparse 3D object detectors. Inspired by the recently proposed works, we further improve our model's efficiency by vertically compressing the image features when serving as attention keys and values, and then we develop two modules to compensate for potential information loss due to feature compression. Experimental evaluation on the widely-used nuScenes 3D object detection benchmark demonstrates that our method outperforms previous approaches across different 3D detection architectures. More importantly, our model is highly efficient. For example, when using $256\times 704$ input images, it achieves 1.5 ms and 2.8 ms latency on NVIDIA 3090 GPU and Horizon Journey-5 computation solutions. Furthermore, WidthFormer also exhibits strong robustness to different degrees of camera perturbations. Our study offers valuable insights into the deployment of BEV transformation methods in real-world, complex road environments. Code is available at https://github.com/ChenhongyiYang/WidthFormer .
High-quality imaging is crucial for ensuring safety supervision and intelligent deployment in fields like transportation and industry. It enables precise and detailed monitoring of operations, facilitating timely detection of potential hazards and efficient management. However, adverse weather conditions, such as atmospheric haziness and precipitation, can have a significant impact on image quality. When the atmosphere contains dense haze or water droplets, the incident light scatters, leading to degraded captured images. This degradation is evident in the form of image blur and reduced contrast, increasing the likelihood of incorrect assessments and interpretations by intelligent imaging systems (IIS). To address the challenge of restoring degraded images in hazy and rainy conditions, this paper proposes a novel multi-view knowledge-guided scene recovery network (termed MvKSR). Specifically, guided filtering is performed on the degraded image to separate high/low-frequency components. Subsequently, an en-decoder-based multi-view feature coarse extraction module (MCE) is used to coarsely extract features from different views of the degraded image. The multi-view feature fine fusion module (MFF) will learn and infer the restoration of degraded images through mixed supervision under different views. Additionally, we suggest an atrous residual block to handle global restoration and local repair in hazy/rainy/mixed scenes. Extensive experimental results demonstrate that MvKSR outperforms other state-of-the-art methods in terms of efficiency and stability for restoring degraded scenarios in IIS.
Accurate per-object distance estimation is crucial in safety-critical applications such as autonomous driving, surveillance, and robotics. Existing approaches rely on two scales: local information (i.e., the bounding box proportions) or global information, which encodes the semantics of the scene as well as the spatial relations with neighboring objects. However, these approaches may struggle with long-range objects and in the presence of strong occlusions or unusual visual patterns. In this respect, our work aims to strengthen both local and global cues. Our architecture -- named DistFormer -- builds upon three major components acting jointly: i) a robust context encoder extracting fine-grained per-object representations; ii) a masked encoder-decoder module exploiting self-supervision to promote the learning of useful per-object features; iii) a global refinement module that aggregates object representations and computes a joint, spatially-consistent estimation. To evaluate the effectiveness of DistFormer, we conduct experiments on the standard KITTI dataset and the large-scale NuScenes and MOTSynth datasets. Such datasets cover various indoor/outdoor environments, changing weather conditions, appearances, and camera viewpoints. Our comprehensive analysis shows that DistFormer outperforms existing methods. Moreover, we further delve into its generalization capabilities, showing its regularization benefits in zero-shot synth-to-real transfer.
DETR accomplishes end-to-end object detection through iteratively generating multiple object candidates based on image features and promoting one candidate for each ground-truth object. The traditional training procedure using one-to-one supervision in the original DETR lacks direct supervision for the object detection candidates. We aim at improving the DETR training efficiency by explicitly supervising the candidate generation procedure through mixing one-to-one supervision and one-to-many supervision. Our approach, namely MS-DETR, is simple, and places one-to-many supervision to the object queries of the primary decoder that is used for inference. In comparison to existing DETR variants with one-to-many supervision, such as Group DETR and Hybrid DETR, our approach does not need additional decoder branches or object queries. The object queries of the primary decoder in our approach directly benefit from one-to-many supervision and thus are superior in object candidate prediction. Experimental results show that our approach outperforms related DETR variants, such as DN-DETR, Hybrid DETR, and Group DETR, and the combination with related DETR variants further improves the performance.
Limited by the encoder-decoder architecture, learning-based edge detectors usually have difficulty predicting edge maps that satisfy both correctness and crispness. With the recent success of the diffusion probabilistic model (DPM), we found it is especially suitable for accurate and crisp edge detection since the denoising process is directly applied to the original image size. Therefore, we propose the first diffusion model for the task of general edge detection, which we call DiffusionEdge. To avoid expensive computational resources while retaining the final performance, we apply DPM in the latent space and enable the classic cross-entropy loss which is uncertainty-aware in pixel level to directly optimize the parameters in latent space in a distillation manner. We also adopt a decoupled architecture to speed up the denoising process and propose a corresponding adaptive Fourier filter to adjust the latent features of specific frequencies. With all the technical designs, DiffusionEdge can be stably trained with limited resources, predicting crisp and accurate edge maps with much fewer augmentation strategies. Extensive experiments on four edge detection benchmarks demonstrate the superiority of DiffusionEdge both in correctness and crispness. On the NYUDv2 dataset, compared to the second best, we increase the ODS, OIS (without post-processing) and AC by 30.2%, 28.1% and 65.1%, respectively. Code: https://github.com/GuHuangAI/DiffusionEdge.
In the realm of artificial intelligence, the emergence of foundation models, backed by high computing capabilities and extensive data, has been revolutionary. Segment Anything Model (SAM), built on the Vision Transformer (ViT) model with millions of parameters and vast training dataset SA-1B, excels in various segmentation scenarios relying on its significance of semantic information and generalization ability. Such achievement of visual foundation model stimulates continuous researches on specific downstream tasks in computer vision. The ClassWise-SAM-Adapter (CWSAM) is designed to adapt the high-performing SAM for landcover classification on space-borne Synthetic Aperture Radar (SAR) images. The proposed CWSAM freezes most of SAM's parameters and incorporates lightweight adapters for parameter efficient fine-tuning, and a classwise mask decoder is designed to achieve semantic segmentation task. This adapt-tuning method allows for efficient landcover classification of SAR images, balancing the accuracy with computational demand. In addition, the task specific input module injects low frequency information of SAR images by MLP-based layers to improve the model performance. Compared to conventional state-of-the-art semantic segmentation algorithms by extensive experiments, CWSAM showcases enhanced performance with fewer computing resources, highlighting the potential of leveraging foundational models like SAM for specific downstream tasks in the SAR domain. The source code is available at: https://github.com/xypu98/CWSAM.
We characterize and study zero-shot abstractive summarization in Large Language Models (LLMs) by measuring position bias, which we propose as a general formulation of the more restrictive lead bias phenomenon studied previously in the literature. Position bias captures the tendency of a model unfairly prioritizing information from certain parts of the input text over others, leading to undesirable behavior. Through numerous experiments on four diverse real-world datasets, we study position bias in multiple LLM models such as GPT 3.5-Turbo, Llama-2, and Dolly-v2, as well as state-of-the-art pretrained encoder-decoder abstractive summarization models such as Pegasus and BART. Our findings lead to novel insights and discussion on performance and position bias of models for zero-shot summarization tasks.
Tactics, Techniques, and Procedures (TTPs) outline the methods attackers use to exploit vulnerabilities. The interpretation of TTPs in the MITRE ATT&CK framework can be challenging for cybersecurity practitioners due to presumed expertise and complex dependencies. Meanwhile, advancements with Large Language Models (LLMs) have led to recent surge in studies exploring its uses in cybersecurity operations. It is, however, unclear how LLMs can be used in an efficient and proper way to provide accurate responses for critical domains such as cybersecurity. This leads us to investigate how to better use two types of LLMs: small-scale encoder-only (e.g., RoBERTa) and larger decoder-only (e.g., GPT-3.5) LLMs to comprehend and summarize TTPs with the intended purposes (i.e., tactics) of a cyberattack procedure. This work studies and compares the uses of supervised fine-tuning (SFT) of encoder-only LLMs vs. Retrieval Augmented Generation (RAG) for decoder-only LLMs (without fine-tuning). Both SFT and RAG techniques presumably enhance the LLMs with relevant contexts for each cyberattack procedure. Our studies show decoder-only LLMs with RAG achieves better performance than encoder-only models with SFT, particularly when directly relevant context is extracted by RAG. The decoder-only results could suffer low `Precision' while achieving high `Recall'. Our findings further highlight a counter-intuitive observation that more generic prompts tend to yield better predictions of cyberattack tactics than those that are more specifically tailored.
The generative priors of pre-trained latent diffusion models (DMs) have demonstrated great potential to enhance the visual quality of image super-resolution (SR) results. However, the noise sampling process in DMs introduces randomness in the SR outputs, and the generated contents can differ a lot with different noise samples. The multi-step diffusion process can be accelerated by distilling methods, but the generative capacity is difficult to control. To address these issues, we analyze the respective advantages of DMs and generative adversarial networks (GANs) and propose to partition the generative SR process into two stages, where the DM is employed for reconstructing image structures and the GAN is employed for improving fine-grained details. Specifically, we propose a non-uniform timestep sampling strategy in the first stage. A single timestep sampling is first applied to extract the coarse information from the input image, then a few reverse steps are used to reconstruct the main structures. In the second stage, we finetune the decoder of the pre-trained variational auto-encoder by adversarial GAN training for deterministic detail enhancement. Once trained, our proposed method, namely content consistent super-resolution (CCSR),allows flexible use of different diffusion steps in the inference stage without re-training. Extensive experiments show that with 2 or even 1 diffusion step, CCSR can significantly improve the content consistency of SR outputs while keeping high perceptual quality. Codes and models can be found at \href{https://github.com/csslc/CCSR}{https://github.com/csslc/CCSR}.
Accurate medical image segmentation is essential for clinical quantification, disease diagnosis, treatment planning and many other applications. Both convolution-based and transformer-based u-shaped architectures have made significant success in various medical image segmentation tasks. The former can efficiently learn local information of images while requiring much more image-specific inductive biases inherent to convolution operation. The latter can effectively capture long-range dependency at different feature scales using self-attention, whereas it typically encounters the challenges of quadratic compute and memory requirements with sequence length increasing. To address this problem, through integrating the merits of these two paradigms in a well-designed u-shaped architecture, we propose a hybrid yet effective CNN-Transformer network, named BRAU-Net++, for an accurate medical image segmentation task. Specifically, BRAU-Net++ uses bi-level routing attention as the core building block to design our u-shaped encoder-decoder structure, in which both encoder and decoder are hierarchically constructed, so as to learn global semantic information while reducing computational complexity. Furthermore, this network restructures skip connection by incorporating channel-spatial attention which adopts convolution operations, aiming to minimize local spatial information loss and amplify global dimension-interaction of multi-scale features. Extensive experiments on three public benchmark datasets demonstrate that our proposed approach surpasses other state-of-the-art methods including its baseline: BRAU-Net under almost all evaluation metrics. We achieve the average Dice-Similarity Coefficient (DSC) of 82.47, 90.10, and 92.94 on Synapse multi-organ segmentation, ISIC-2018 Challenge, and CVC-ClinicDB, as well as the mIoU of 84.01 and 88.17 on ISIC-2018 Challenge and CVC-ClinicDB, respectively.
We introduce \textit{GraphGPT}, a novel model for Graph learning by self-supervised Generative Pre-training Transformers. Our model transforms each graph or sampled subgraph into a sequence of tokens representing the node, edge and attributes reversibly using the Eulerian path first. Then we feed the tokens into a standard transformer decoder and pre-train it with the next-token-prediction (NTP) task. Lastly, we fine-tune the GraphGPT model with the supervised tasks. This intuitive, yet effective model achieves superior or close results to the state-of-the-art methods for the graph-, edge- and node-level tasks on the large scale molecular dataset PCQM4Mv2, the protein-protein association dataset ogbl-ppa and the ogbn-proteins dataset from the Open Graph Benchmark (OGB). Furthermore, the generative pre-training enables us to train GraphGPT up to 400M+ parameters with consistently increasing performance, which is beyond the capability of GNNs and previous graph transformers. The source code and pre-trained checkpoints will be released soon\footnote{\url{https://github.com/alibaba/graph-gpt}} to pave the way for the graph foundation model research, and also to assist the scientific discovery in pharmaceutical, chemistry, material and bio-informatics domains, etc.
In this paper, we propose a novel method for joint entity and relation extraction from unstructured text by framing it as a conditional sequence generation problem. In contrast to conventional generative information extraction models that are left-to-right token-level generators, our approach is \textit{span-based}. It generates a linearized graph where nodes represent text spans and edges represent relation triplets. Our method employs a transformer encoder-decoder architecture with pointing mechanism on a dynamic vocabulary of spans and relation types. Our model can capture the structural characteristics and boundaries of entities and relations through span representations while simultaneously grounding the generated output in the original text thanks to the pointing mechanism. Evaluation on benchmark datasets validates the effectiveness of our approach, demonstrating competitive results. Code is available at https://github.com/urchade/ATG.