Scene flow estimation determines a scene's 3D motion field, by predicting the motion of points in the scene, especially for aiding tasks in autonomous driving. Many networks with large-scale point clouds as input use voxelization to create a pseudo-image for real-time running. However, the voxelization process often results in the loss of point-specific features. This gives rise to a challenge in recovering those features for scene flow tasks. Our paper introduces DeFlow which enables a transition from voxel-based features to point features using Gated Recurrent Unit (GRU) refinement. To further enhance scene flow estimation performance, we formulate a novel loss function that accounts for the data imbalance between static and dynamic points. Evaluations on the Argoverse 2 scene flow task reveal that DeFlow achieves state-of-the-art results on large-scale point cloud data, demonstrating that our network has better performance and efficiency compared to others. The code is open-sourced at https://github.com/KTH-RPL/deflow.
Videos captured from multiple viewpoints can help in perceiving the 3D structure of the world and benefit computer vision tasks such as action recognition, tracking, etc. In this paper, we present a method for self-supervised learning from synchronized multi-view videos. We use a cross-view reconstruction task to inject geometry information in the model. Our approach is based on the masked autoencoder (MAE) framework. In addition to the same-view decoder, we introduce a separate cross-view decoder which leverages cross-attention mechanism to reconstruct a target viewpoint video using a video from source viewpoint, to help representations robust to viewpoint changes. For videos, static regions can be reconstructed trivially which hinders learning meaningful representations. To tackle this, we introduce a motion-weighted reconstruction loss which improves temporal modeling. We report state-of-the-art results on the NTU-60, NTU-120 and ETRI datasets, as well as in the transfer learning setting on NUCLA, PKU-MMD-II and ROCOG-v2 datasets, demonstrating the robustness of our approach. Code will be made available.
The rapid development of generative diffusion models has significantly advanced the field of style transfer. However, most current style transfer methods based on diffusion models typically involve a slow iterative optimization process, e.g., model fine-tuning and textual inversion of style concept. In this paper, we introduce FreeStyle, an innovative style transfer method built upon a pre-trained large diffusion model, requiring no further optimization. Besides, our method enables style transfer only through a text description of the desired style, eliminating the necessity of style images. Specifically, we propose a dual-stream encoder and single-stream decoder architecture, replacing the conventional U-Net in diffusion models. In the dual-stream encoder, two distinct branches take the content image and style text prompt as inputs, achieving content and style decoupling. In the decoder, we further modulate features from the dual streams based on a given content image and the corresponding style text prompt for precise style transfer. Our experimental results demonstrate high-quality synthesis and fidelity of our method across various content images and style text prompts. Compared with state-of-the-art methods that require training, our FreeStyle approach notably reduces the computational burden by thousands of iterations, while achieving comparable or superior performance across multiple evaluation metrics including CLIP Aesthetic Score, CLIP Score, and Preference. We have released the code at: https://github.com/FreeStyleFreeLunch/FreeStyle.
Detecting glass regions is a challenging task due to the ambiguity of their transparency and reflection properties. These transparent glasses share the visual appearance of both transmitted arbitrary background scenes and reflected objects, thus having no fixed patterns.Recent visual foundation models, which are trained on vast amounts of data, have manifested stunning performance in terms of image perception and image generation. To segment glass surfaces with higher accuracy, we make full use of two visual foundation models: Segment Anything (SAM) and Stable Diffusion.Specifically, we devise a simple glass surface segmentor named GEM, which only consists of a SAM backbone, a simple feature pyramid, a discerning query selection module, and a mask decoder. The discerning query selection can adaptively identify glass surface features, assigning them as initialized queries in the mask decoder. We also propose a Synthetic but photorealistic large-scale Glass Surface Detection dataset dubbed S-GSD via diffusion model with four different scales, which contain 1x, 5x, 10x, and 20x of the original real data size. This dataset is a feasible source for transfer learning. The scale of synthetic data has positive impacts on transfer learning, while the improvement will gradually saturate as the amount of data increases. Extensive experiments demonstrate that GEM achieves a new state-of-the-art on the GSD-S validation set (IoU +2.1%). Codes and datasets are available at: https://github.com/isbrycee/GEM-Glass-Segmentor.
BERT (Bidirectional Encoder Representations from Transformers) has revolutionized the field of natural language processing through its exceptional performance on numerous tasks. Yet, the majority of researchers have mainly concentrated on enhancements related to the model structure, such as relative position embedding and more efficient attention mechanisms. Others have delved into pretraining tricks associated with Masked Language Modeling, including whole word masking. DeBERTa introduced an enhanced decoder adapted for BERT's encoder model for pretraining, proving to be highly effective. We argue that the design and research around enhanced masked language modeling decoders have been underappreciated. In this paper, we propose several designs of enhanced decoders and introduce BPDec (BERT Pretraining Decoder), a novel method for modeling training. Typically, a pretrained BERT model is fine-tuned for specific Natural Language Understanding (NLU) tasks. In our approach, we utilize the original BERT model as the encoder, making only changes to the decoder without altering the encoder. This approach does not necessitate extensive modifications to the encoder architecture and can be seamlessly integrated into existing fine-tuning pipelines and services, offering an efficient and effective enhancement strategy. Compared to other methods, while we also incur a moderate training cost for the decoder during the pretraining process, our approach does not introduce additional training costs during the fine-tuning phase. We test multiple enhanced decoder structures after pretraining and evaluate their performance on the GLUE tasks and SQuAD tasks. Our results demonstrate that BPDec, having only undergone subtle refinements to the model structure during pretraining, significantly enhances model performance without escalating the finetuning cost, inference time and serving budget.
Improving the efficiency of state-of-the-art methods in semantic segmentation requires overcoming the increasing computational cost as well as issues such as fusing semantic information from global and local contexts. Based on the recent success and problems that convolutional neural networks (CNNs) encounter in semantic segmentation, this research proposes an encoder-decoder architecture with a unique efficient residual network, Efficient-ResNet. Attention-boosting gates (AbGs) and attention-boosting modules (AbMs) are deployed by aiming to fuse the equivariant and feature-based semantic information with the equivalent sizes of the output of global context of the efficient residual network in the encoder. Respectively, the decoder network is developed with the additional attention-fusion networks (AfNs) inspired by AbM. AfNs are designed to improve the efficiency in the one-to-one conversion of the semantic information by deploying additional convolution layers in the decoder part. Our network is tested on the challenging CamVid and Cityscapes datasets, and the proposed methods reveal significant improvements on the residual networks. To the best of our knowledge, the developed network, SERNet-Former, achieves state-of-the-art results (84.62 % mean IoU) on CamVid dataset and challenging results (87.35 % mean IoU) on Cityscapes validation dataset.
Task-oriented dialog systems have witnessed substantial progress due to conversational pre-training techniques. Yet, two significant challenges persist. First, most systems primarily utilize the latest turn's state label for the generator. This practice overlooks the comprehensive value of state labels in boosting the model's understanding for future generations. Second, an overreliance on generated policy often leads to error accumulation, resulting in suboptimal responses when adhering to incorrect actions. To combat these challenges, we propose turn-level multi-task objectives for the encoder. With the guidance of essential information from labeled intermediate states, we establish a more robust representation for both understanding and generation. For the decoder, we introduce an action tree-based scheduled sampling technique. Specifically, we model the hierarchical policy as trees and utilize the similarity between trees to sample negative policy based on scheduled sampling, hoping the model to generate invariant responses under perturbations. This method simulates potential pitfalls by sampling similar negative policy, bridging the gap between task-oriented dialog training and inference. Among methods without continual pre-training, our approach achieved state-of-the-art (SOTA) performance on the MultiWOZ dataset series and was also competitive with pre-trained SOTA methods.
Dysarthric speech reconstruction (DSR) systems aim to automatically convert dysarthric speech into normal-sounding speech. The technology eases communication with speakers affected by the neuromotor disorder and enhances their social inclusion. NED-based (Neural Encoder-Decoder) systems have significantly improved the intelligibility of the reconstructed speech as compared with GAN-based (Generative Adversarial Network) approaches, but the approach is still limited by training inefficiency caused by the cascaded pipeline and auxiliary tasks of the content encoder, which may in turn affect the quality of reconstruction. Inspired by self-supervised speech representation learning and discrete speech units, we propose a Unit-DSR system, which harnesses the powerful domain-adaptation capacity of HuBERT for training efficiency improvement and utilizes speech units to constrain the dysarthric content restoration in a discrete linguistic space. Compared with NED approaches, the Unit-DSR system only consists of a speech unit normalizer and a Unit HiFi-GAN vocoder, which is considerably simpler without cascaded sub-modules or auxiliary tasks. Results on the UASpeech corpus indicate that Unit-DSR outperforms competitive baselines in terms of content restoration, reaching a 28.2% relative average word error rate reduction when compared to original dysarthric speech, and shows robustness against speed perturbation and noise.
Pre-trained language models based on masked language modeling (MLM) excel in natural language understanding (NLU) tasks. While fine-tuned MLM-based encoders consistently outperform causal language modeling decoders of comparable size, recent decoder-only large language models (LLMs) perform on par with smaller MLM-based encoders. Although their performance improves with scale, LLMs fall short of achieving state-of-the-art results in information extraction (IE) tasks, many of which are formulated as sequence labeling (SL). We hypothesize that LLMs' poor SL performance stems from causal masking, which prevents the model from attending to tokens on the right of the current token. Yet, how exactly and to what extent LLMs' performance on SL can be improved remains unclear. We explore techniques for improving the SL performance of open LLMs on IE tasks by applying layer-wise removal of the causal mask (CM) during LLM fine-tuning. This approach yields performance gains competitive with state-of-the-art SL models, matching or outperforming the results of CM removal from all blocks. Our findings hold for diverse SL tasks, demonstrating that open LLMs with layer-dependent CM removal outperform strong MLM-based encoders and even instruction-tuned LLMs.
Scene-based spatial audio formats, such as Ambisonics, are playback system agnostic and may therefore be favoured for delivering immersive audio experiences to a wide range of (potentially unknown) devices. The number of channels required to deliver high spatial resolution Ambisonic audio, however, can be prohibitive for low-bandwidth applications. Therefore, this paper proposes a compression codec, which is based upon the parametric higher-order Directional Audio Coding (HO-DirAC) model. The encoder downmixes the higher-order Ambisonic (HOA) input audio into a reduced number of signals, which are accompanied by perceptually-motivated scene parameters. The downmixed audio is coded using a perceptual audio coder, whereas the parameters are grouped into perceptual bands, quantized, and downsampled. On the decoder side, low Ambisonic orders are fully recovered. Not fully recoverable HOA components are synthesized according to the parameters. The results of a listening test indicate that the proposed parametric spatial audio codec can improve the adopted perceptual audio coder, especially at low to medium-high bitrates, when applied to fifth-order HOA signals.
Recent TTS models with decoder-only Transformer architecture, such as SPEAR-TTS and VALL-E, achieve impressive naturalness and demonstrate the ability for zero-shot adaptation given a speech prompt. However, such decoder-only TTS models lack monotonic alignment constraints, sometimes leading to hallucination issues such as mispronunciation, word skipping and repeating. To address this limitation, we propose VALL-T, a generative Transducer model that introduces shifting relative position embeddings for input phoneme sequence, explicitly indicating the monotonic generation process while maintaining the architecture of decoder-only Transformer. Consequently, VALL-T retains the capability of prompt-based zero-shot adaptation and demonstrates better robustness against hallucinations with a relative reduction of 28.3% in the word error rate.
Hierarchical topic modeling aims to discover latent topics from a corpus and organize them into a hierarchy to understand documents with desirable semantic granularity. However, existing work struggles with producing topic hierarchies of low affinity, rationality, and diversity, which hampers document understanding. To overcome these challenges, we in this paper propose Transport Plan and Context-aware Hierarchical Topic Model (TraCo). Instead of early simple topic dependencies, we propose a transport plan dependency method. It constrains dependencies to ensure their sparsity and balance, and also regularizes topic hierarchy building with them. This improves affinity and diversity of hierarchies. We further propose a context-aware disentangled decoder. Rather than previously entangled decoding, it distributes different semantic granularity to topics at different levels by disentangled decoding. This facilitates the rationality of hierarchies. Experiments on benchmark datasets demonstrate that our method surpasses state-of-the-art baselines, effectively improving the affinity, rationality, and diversity of hierarchical topic modeling with better performance on downstream tasks.
Inverse molecular design with diffusion models holds great potential for advancements in material and drug discovery. Despite success in unconditional molecular generation, integrating multiple properties such as synthetic score and gas permeability as condition constraints into diffusion models remains unexplored. We present the Graph Diffusion Transformer (Graph DiT) for multi-conditional molecular generation. Graph DiT integrates an encoder to learn numerical and categorical property representations with the Transformer-based denoiser. Unlike previous graph diffusion models that add noise separately on the atoms and bonds in the forward diffusion process, Graph DiT is trained with a novel graph-dependent noise model for accurate estimation of graph-related noise in molecules. We extensively validate Graph DiT for multi-conditional polymer and small molecule generation. Results demonstrate the superiority of Graph DiT across nine metrics from distribution learning to condition control for molecular properties. A polymer inverse design task for gas separation with feedback from domain experts further demonstrates its practical utility.
In this work, we re-examine inter-patch dependencies in the decoding mechanism of masked autoencoders (MAE). We decompose this decoding mechanism for masked patch reconstruction in MAE into self-attention and cross-attention. Our investigations suggest that self-attention between mask patches is not essential for learning good representations. To this end, we propose a novel pretraining framework: Cross-Attention Masked Autoencoders (CrossMAE). CrossMAE's decoder leverages only cross-attention between masked and visible tokens, with no degradation in downstream performance. This design also enables decoding only a small subset of mask tokens, boosting efficiency. Furthermore, each decoder block can now leverage different encoder features, resulting in improved representation learning. CrossMAE matches MAE in performance with 2.5 to 3.7$\times$ less decoding compute. It also surpasses MAE on ImageNet classification and COCO instance segmentation under the same compute. Code and models: https://crossmae.github.io
This paper introduces a novel approach called "friendly attack" aimed at enhancing the performance of error correction channel codes. Inspired by the concept of adversarial attacks, our method leverages the idea of introducing slight perturbations to the neural network input, resulting in a substantial impact on the network's performance. By introducing small perturbations to fixed-point modulated codewords before transmission, we effectively improve the decoder's performance without violating the input power constraint. The perturbation design is accomplished by a modified iterative fast gradient method. This study investigates various decoder architectures suitable for computing gradients to obtain the desired perturbations. Specifically, we consider belief propagation (BP) for LDPC codes; the error correcting code transformer, BP and neural BP (NBP) for polar codes, and neural BCJR for convolutional codes. We demonstrate that the proposed friendly attack method can improve the reliability across different channels, modulations, codes, and decoders. This method allows us to increase the reliability of communication with a legacy receiver by simply modifying the transmitted codeword appropriately.
Despite the recent emergence of video captioning models, how to generate the text description with specific entity names and fine-grained actions is far from being solved, which however has great applications such as basketball live text broadcast. In this paper, a new multimodal knowledge graph supported basketball benchmark for video captioning is proposed. Specifically, we construct a multimodal basketball game knowledge graph (KG_NBA_2022) to provide additional knowledge beyond videos. Then, a multimodal basketball game video captioning (VC_NBA_2022) dataset that contains 9 types of fine-grained shooting events and 286 players' knowledge (i.e., images and names) is constructed based on KG_NBA_2022. We develop a knowledge guided entity-aware video captioning network (KEANet) based on a candidate player list in encoder-decoder form for basketball live text broadcast. The temporal contextual information in video is encoded by introducing the bi-directional GRU (Bi-GRU) module. And the entity-aware module is designed to model the relationships among the players and highlight the key players. Extensive experiments on multiple sports benchmarks demonstrate that KEANet effectively leverages extera knowledge and outperforms advanced video captioning models. The proposed dataset and corresponding codes will be publicly available soon
The recent advances in natural language processing have predominantly favored well-resourced English-centric models, resulting in a significant gap with low-resource languages. In this work, we introduce the language model TURNA, which is developed for the low-resource language Turkish and is capable of both natural language understanding and generation tasks. TURNA is pretrained with an encoder-decoder architecture based on the unified framework UL2 with a diverse corpus that we specifically curated for this purpose. We evaluated TURNA with three generation tasks and five understanding tasks for Turkish. The results show that TURNA outperforms several multilingual models in both understanding and generation tasks, and competes with monolingual Turkish models in understanding tasks. TURNA is made available at https://huggingface.co/boun-tabi-LMG/TURNA .
Despite the advancement in computational modeling towards brain tumor segmentation, of which several models have been developed, it is evident from the computational complexity of existing models that performance and efficiency under clinical application scenarios are still limited. Therefore, this paper proposes a tumor segmentation framework. It includes a novel shallow encoder and decoder network named SEDNet for brain tumor segmentation. The highlights of SEDNet include sufficiency in hierarchical convolutional downsampling and selective skip mechanism for cost-efficient and effective brain tumor semantic segmentation, among other features. The preprocessor and optimization function approaches are devised to minimize the uncertainty in feature learning impacted by nontumor slices or empty masks with corresponding brain slices and address class imbalances as well as boundary irregularities of tumors, respectively. Through experiments, SEDNet achieved impressive dice and Hausdorff scores of 0.9308 %, 0.9451 %, and 0.9026 %, and 0.7040 mm, 1.2866 mm, and 0.7762 mm for the non-enhancing tumor core (NTC), peritumoral edema (ED), and enhancing tumor (ET), respectively. This is one of the few works to report segmentation performance on NTC. Furthermore, through transfer learning with initialized SEDNet pre-trained weights, termed SEDNetX, a performance increase is observed. The dice and Hausdorff scores recorded are 0.9336%, 0.9478%, 0.9061%, 0.6983 mm, 1.2691 mm, and 0.7711 mm for NTC, ED, and ET, respectively. With about 1.3 million parameters and impressive performance in comparison to the state-of-the-art, SEDNet(X) is shown to be computationally efficient for real-time clinical diagnosis. The code is available on Github .
Pre-training large language models is known to be extremely resource intensive and often times inefficient, under-utilizing the information encapsulated in the training text sequences. In this paper, we present SpacTor, a new training procedure consisting of (1) a hybrid objective combining span corruption (SC) and token replacement detection (RTD), and (2) a two-stage curriculum that optimizes the hybrid objective over the initial $\tau$ iterations, then transitions to standard SC loss. We show empirically that the effectiveness of the hybrid objective is tied to the two-stage pre-training schedule, and provide extensive analysis on why this is the case. In our experiments with encoder-decoder architectures (T5) on a variety of NLP tasks, SpacTor-T5 yields the same downstream performance as standard SC pre-training, while enabling a 50% reduction in pre-training iterations and 40% reduction in total FLOPs. Alternatively, given the same amount of computing budget, we find that SpacTor results in significantly improved downstream benchmark performance.
The Segment Anything Model (SAM) has exhibited outstanding performance in various image segmentation tasks. Despite being trained with over a billion masks, SAM faces challenges in mask prediction quality in numerous scenarios, especially in real-world contexts. In this paper, we introduce a novel prompt-driven adapter into SAM, namely Prompt Adapter Segment Anything Model (PA-SAM), aiming to enhance the segmentation mask quality of the original SAM. By exclusively training the prompt adapter, PA-SAM extracts detailed information from images and optimizes the mask decoder feature at both sparse and dense prompt levels, improving the segmentation performance of SAM to produce high-quality masks. Experimental results demonstrate that our PA-SAM outperforms other SAM-based methods in high-quality, zero-shot, and open-set segmentation. We're making the source code and models available at https://github.com/xzz2/pa-sam.