Masked graph modeling excels in the self-supervised representation learning of molecular graphs. Scrutinizing previous studies, we can reveal a common scheme consisting of three key components: (1) graph tokenizer, which breaks a molecular graph into smaller fragments (i.e., subgraphs) and converts them into tokens; (2) graph masking, which corrupts the graph with masks; (3) graph autoencoder, which first applies an encoder on the masked graph to generate the representations, and then employs a decoder on the representations to recover the tokens of the original graph. However, the previous MGM studies focus extensively on graph masking and encoder, while there is limited understanding of tokenizer and decoder. To bridge the gap, we first summarize popular molecule tokenizers at the granularity of node, edge, motif, and Graph Neural Networks (GNNs), and then examine their roles as the MGM's reconstruction targets. Further, we explore the potential of adopting an expressive decoder in MGM. Our results show that a subgraph-level tokenizer and a sufficiently expressive decoder with remask decoding have a large impact on the encoder's representation learning. Finally, we propose a novel MGM method SimSGT, featuring a Simple GNN-based Tokenizer (SGT) and an effective decoding strategy. We empirically validate that our method outperforms the existing molecule self-supervised learning methods. Our codes and checkpoints are available at https://github.com/syr-cn/SimSGT.
Depth completion is a critical task for handling depth images with missing pixels, which can negatively impact further applications. Recent approaches have utilized Convolutional Neural Networks (CNNs) to reconstruct depth images with the assistance of color images. However, vanilla convolution has non-negligible drawbacks in handling missing pixels. To solve this problem, we propose a new model for depth completion based on an encoder-decoder structure. Our model introduces two key components: the Mask-adaptive Gated Convolution (MagaConv) architecture and the Bi-directional Progressive Fusion (BP-Fusion) module. The MagaConv architecture is designed to acquire precise depth features by modulating convolution operations with iteratively updated masks, while the BP-Fusion module progressively integrates depth and color features, utilizing consecutive bi-directional fusion structures in a global perspective. Extensive experiments on popular benchmarks, including NYU-Depth V2, DIML, and SUN RGB-D, demonstrate the superiority of our model over state-of-the-art methods. We achieved remarkable performance in completing depth maps and outperformed existing approaches in terms of accuracy and reliability.
3D landmark detection plays a pivotal role in various applications such as 3D registration, pose estimation, and virtual try-on. While considerable success has been achieved in 2D human landmark detection or pose estimation, there is a notable scarcity of reported works on landmark detection in unordered 3D point clouds. This paper introduces a novel challenge, namely 3D landmark detection on human point clouds, presenting two primary contributions. Firstly, we establish a comprehensive human point cloud dataset, named HPoint103, designed to support the 3D landmark detection community. This dataset comprises 103 human point clouds created with commercial software and actors, each manually annotated with 11 stable landmarks. Secondly, we propose a Dual Cascade Point Transformer (D-CPT) model for precise point-based landmark detection. D-CPT gradually refines the landmarks through cascade Transformer decoder layers across the entire point cloud stream, simultaneously enhancing landmark coordinates with a RefineNet over local regions. Comparative evaluations with popular point-based methods on HPoint103 and the public dataset DHP19 demonstrate the dramatic outperformance of our D-CPT. Additionally, the integration of our RefineNet into existing methods consistently improves performance.
Natural Language Video Localization (NLVL), grounding phrases from natural language descriptions to corresponding video segments, is a complex yet critical task in video understanding. Despite ongoing advancements, many existing solutions lack the capability to globally capture temporal dynamics of the video data. In this study, we present a novel approach to NLVL that aims to address this issue. Our method involves the direct generation of a global 2D temporal map via a conditional denoising diffusion process, based on the input video and language query. The main challenges are the inherent sparsity and discontinuity of a 2D temporal map in devising the diffusion decoder. To address these challenges, we introduce a multi-scale technique and develop an innovative diffusion decoder. Our approach effectively encapsulates the interaction between the query and video data across various time scales. Experiments on the Charades and DiDeMo datasets underscore the potency of our design.
In recent years, finding an effective and efficient strategy for exploiting spatial and temporal information has been a hot research topic in video saliency prediction (VSP). With the emergence of spatio-temporal transformers, the weakness of the prior strategies, e.g., 3D convolutional networks and LSTM-based networks, for capturing long-range dependencies has been effectively compensated. While VSP has drawn benefits from spatio-temporal transformers, finding the most effective way for aggregating temporal features is still challenging. To address this concern, we propose a transformer-based video saliency prediction approach with high temporal dimension decoding network (THTD-Net). This strategy accounts for the lack of complex hierarchical interactions between features that are extracted from the transformer-based spatio-temporal encoder: in particular, it does not require multiple decoders and aims at gradually reducing temporal features' dimensions in the decoder. This decoder-based architecture yields comparable performance to multi-branch and over-complicated models on common benchmarks such as DHF1K, UCF-sports and Hollywood-2.
Data protection methods like cryptography, despite being effective, inadvertently signal the presence of secret communication, thereby drawing undue attention. Here, we introduce an optical information hiding camera integrated with an electronic decoder, optimized jointly through deep learning. This information hiding-decoding system employs a diffractive optical processor as its front-end, which transforms and hides input images in the form of ordinary-looking patterns that deceive/mislead human observers. This information hiding transformation is valid for infinitely many combinations of secret messages, all of which are transformed into ordinary-looking output patterns, achieved all-optically through passive light-matter interactions within the optical processor. By processing these ordinary-looking output images, a jointly-trained electronic decoder neural network accurately reconstructs the original information hidden within the deceptive output pattern. We numerically demonstrated our approach by designing an information hiding diffractive camera along with a jointly-optimized convolutional decoder neural network. The efficacy of this system was demonstrated under various lighting conditions and noise levels, showing its robustness. We further extended this information hiding camera to multi-spectral operation, allowing the concealment and decoding of multiple images at different wavelengths, all performed simultaneously in a single feed-forward operation. The feasibility of our framework was also demonstrated experimentally using THz radiation. This optical encoder-electronic decoder-based co-design provides a novel information hiding camera interface that is both high-speed and energy-efficient, offering an intriguing solution for visual information security.
Unlike single image task, stereo image enhancement can use another view information, and its key stage is how to perform cross-view feature interaction to extract useful information from another view. However, complex noise in low-light image and its impact on subsequent feature encoding and interaction are ignored by the existing methods. In this paper, a method is proposed to perform enhancement and de-noising simultaneously. First, to reduce unwanted noise interference, a low-frequency information enhanced module (IEM) is proposed to suppress noise and produce a new image space. Additionally, a cross-channel and spatial context information mining module (CSM) is proposed to encode long-range spatial dependencies and to enhance inter-channel feature interaction. Relying on CSM, an encoder-decoder structure is constructed, incorporating cross-view and cross-scale feature interactions to perform enhancement in the new image space. Finally, the network is trained with the constraints of both spatial and frequency domain losses. Extensive experiments on both synthesized and real datasets show that our method obtains better detail recovery and noise removal compared with state-of-the-art methods. In addition, a real stereo image enhancement dataset is captured with stereo camera ZED2. The code and dataset are publicly available at: https://www.github.com/noportraits/LFENet.
Jewelry recognition is a complex task due to the different styles and designs of accessories. Precise descriptions of the various accessories is something that today can only be achieved by experts in the field of jewelry. In this work, we propose an approach for jewelry recognition using computer vision techniques and image captioning, trying to simulate this expert human behavior of analyzing accessories. The proposed methodology consist on using different image captioning models to detect the jewels from an image and generate a natural language description of the accessory. Then, this description is also utilized to classify the accessories at different levels of detail. The generated caption includes details such as the type of jewel, color, material, and design. To demonstrate the effectiveness of the proposed method in accurately recognizing different types of jewels, a dataset consisting of images of accessories belonging to jewelry stores in C\'ordoba (Spain) has been created. After testing the different image captioning architectures designed, the final model achieves a captioning accuracy of 95\%. The proposed methodology has the potential to be used in various applications such as jewelry e-commerce, inventory management or automatic jewels recognition to analyze people's tastes and social status.
Addition is perhaps one of the simplest arithmetic tasks one can think of and is usually performed using the carrying over algorithm. This algorithm consists of two tasks: adding digits in the same position and carrying over a one whenever necessary. We study how transformer models implement this algorithm and how the two aforementioned tasks are allocated to different parts of the network. We first focus on two-layer encoder-only models and show that the carrying over algorithm is implemented in a modular fashion. The first layer is mostly responsible for adding digits in the same position. The second layer first decides, in the attention, which positions need a carried one or not, and then performs the carrying of the one in the final MLP. We provide a simple way of precisely identifying which neurons are responsible for that task. This implementation of the carrying over algorithm occurs across a range of hyperparameters for two as well as three-layer models. For small decoder-only models, we observe the same implementation and provide suggestive evidence for its existence in three 7B large language models.
The ever-increasing volume of paper submissions makes it difficult to stay informed about the latest state-of-the-art research. To address this challenge, we introduce LEGOBench, a benchmark for evaluating systems that generate scientific leaderboards. LEGOBench is curated from 22 years of preprint submission data on arXiv and more than 11k machine learning leaderboards on the PapersWithCode portal. We present four graph-based and two language model-based leaderboard generation task configurations. We evaluate popular encoder-only scientific language models as well as decoder-only large language models across these task configurations. State-of-the-art models showcase significant performance gaps in automatic leaderboard generation on LEGOBench. The code is available on GitHub ( https://github.com/lingo-iitgn/LEGOBench ) and the dataset is hosted on OSF ( https://osf.io/9v2py/?view_only=6f91b0b510df498ba01595f8f278f94c ).
This paper introduces a novel pipeline to reconstruct the geometry of interacting multi-person in clothing on a globally coherent scene space from a single image. The main challenge arises from the occlusion: a part of a human body is not visible from a single view due to the occlusion by others or the self, which introduces missing geometry and physical implausibility (e.g., penetration). We overcome this challenge by utilizing two human priors for complete 3D geometry and surface contacts. For the geometry prior, an encoder learns to regress the image of a person with missing body parts to the latent vectors; a decoder decodes these vectors to produce 3D features of the associated geometry; and an implicit network combines these features with a surface normal map to reconstruct a complete and detailed 3D humans. For the contact prior, we develop an image-space contact detector that outputs a probability distribution of surface contacts between people in 3D. We use these priors to globally refine the body poses, enabling the penetration-free and accurate reconstruction of interacting multi-person in clothing on the scene space. The results demonstrate that our method is complete, globally coherent, and physically plausible compared to existing methods.
This paper describes the use of connectionist techniques in phonetic speech recognition with strong latency constraints. The constraints are imposed by the task of deriving the lip movements of a synthetic face in real time from the speech signal, by feeding the phonetic string into an articulatory synthesiser. Particular attention has been paid to analysing the interaction between the time evolution model learnt by the multi-layer perceptrons and the transition model imposed by the Viterbi decoder, in different latency conditions. Two experiments were conducted in which the time dependencies in the language model (LM) were controlled by a parameter. The results show a strong interaction between the three factors involved, namely the neural network topology, the length of time dependencies in the LM and the decoder latency.
Transformers are considered conceptually different from the previous generation of state-of-the-art NLP models - recurrent neural networks (RNNs). In this work, we demonstrate that decoder-only transformers can in fact be conceptualized as unbounded multi-state RNNs - an RNN variant with unlimited hidden state size. We further show that transformers can be converted into $\textit{bounded}$ multi-state RNNs by fixing the size of their hidden state, effectively compressing their key-value cache. We introduce a novel, training-free compression policy - $\textbf{T}$oken $\textbf{O}$mission $\textbf{V}$ia $\textbf{A}$ttention (TOVA). Our experiments with four long range tasks and several LLMs show that TOVA outperforms several baseline compression policies. Particularly, our results are nearly on par with the full model, using in some cases only $\frac{1}{8}$ of the original cache size, which translates to 4.8X higher throughput. Our results shed light on the connection between transformers and RNNs, and help mitigate one of LLMs' most painful computational bottlenecks - the size of their key-value cache. We publicly release our code at https://github.com/schwartz-lab-NLP/TOVA
We develop and evaluate two novel purpose-built deep learning (DL) models for synthesis of the arterial blood pressure (ABP) waveform in a cuff-less manner, using a single-site photoplethysmography (PPG) signal. We train and evaluate our DL models on the data of 209 subjects from the public UCI dataset on cuff-less blood pressure (CLBP) estimation. Our transformer model consists of an encoder-decoder pair that incorporates positional encoding, multi-head attention, layer normalization, and dropout techniques for ABP waveform synthesis. Secondly, under our frequency-domain (FD) learning approach, we first obtain the discrete cosine transform (DCT) coefficients of the PPG and ABP signals, and then learn a linear/non-linear (L/NL) regression between them. The transformer model (FD L/NL model) synthesizes the ABP waveform with a mean absolute error (MAE) of 3.01 (4.23). Further, the synthesis of ABP waveform also allows us to estimate the systolic blood pressure (SBP) and diastolic blood pressure (DBP) values. To this end, the transformer model reports an MAE of 3.77 mmHg and 2.69 mmHg, for SBP and DBP, respectively. On the other hand, the FD L/NL method reports an MAE of 4.37 mmHg and 3.91 mmHg, for SBP and DBP, respectively. Both methods fulfill the AAMI criterion. As for the BHS criterion, our transformer model (FD L/NL regression model) achieves grade A (grade B).
Transformer based Large Language Models (LLMs) have been widely used in many fields, and the efficiency of LLM inference becomes hot topic in real applications. However, LLMs are usually complicatedly designed in model structure with massive operations and perform inference in the auto-regressive mode, making it a challenging task to design a system with high efficiency. In this paper, we propose an efficient LLM inference solution with low latency and high throughput. Firstly, we simplify the LLM decoder layer by fusing data movement and element-wise operations to reduce the memory access frequency and lower system latency. We also propose a segment KV cache policy to keep key/value of the request and response tokens in separate physical memory for effective device memory management, helping enlarge the runtime batch size and improve system throughput. A customized Scaled-Dot-Product-Attention kernel is designed to match our fusion policy based on the segment KV cache solution. We implement our LLM inference solution on Intel GPU and publish it publicly. Compared with the standard HuggingFace implementation, the proposed solution achieves up to 7x lower token latency and 27x higher throughput for some popular LLMs on Intel GPU.
Traffic forecasting is the foundation for intelligent transportation systems. Spatiotemporal graph neural networks have demonstrated state-of-the-art performance in traffic forecasting. However, these methods do not explicitly model some of the natural characteristics in traffic data, such as the multiscale structure that encompasses spatial and temporal variations at different levels of granularity or scale. To that end, we propose a Wavelet-Inspired Graph Convolutional Recurrent Network (WavGCRN) which combines multiscale analysis (MSA)-based method with Deep Learning (DL)-based method. In WavGCRN, the traffic data is decomposed into time-frequency components with Discrete Wavelet Transformation (DWT), constructing a multi-stream input structure; then Graph Convolutional Recurrent networks (GCRNs) are employed as encoders for each stream, extracting spatiotemporal features in different scales; and finally the learnable Inversed DWT and GCRN are combined as the decoder, fusing the information from all streams for traffic metrics reconstruction and prediction. Furthermore, road-network-informed graphs and data-driven graph learning are combined to accurately capture spatial correlation. The proposed method can offer well-defined interpretability, powerful learning capability, and competitive forecasting performance on real-world traffic data sets.
Counterfactual inference aims to answer retrospective "what if" questions and thus belongs to the most fine-grained type of inference in Pearl's causality ladder. Existing methods for counterfactual inference with continuous outcomes aim at point identification and thus make strong and unnatural assumptions about the underlying structural causal model. In this paper, we relax these assumptions and aim at partial counterfactual identification of continuous outcomes, i.e., when the counterfactual query resides in an ignorance interval with informative bounds. We prove that, in general, the ignorance interval of the counterfactual queries has non-informative bounds, already when functions of structural causal models are continuously differentiable. As a remedy, we propose a novel sensitivity model called Curvature Sensitivity Model. This allows us to obtain informative bounds by bounding the curvature of level sets of the functions. We further show that existing point counterfactual identification methods are special cases of our Curvature Sensitivity Model when the bound of the curvature is set to zero. We then propose an implementation of our Curvature Sensitivity Model in the form of a novel deep generative model, which we call Augmented Pseudo-Invertible Decoder. Our implementation employs (i) residual normalizing flows with (ii) variational augmentations. We empirically demonstrate the effectiveness of our Augmented Pseudo-Invertible Decoder. To the best of our knowledge, ours is the first partial identification model for Markovian structural causal models with continuous outcomes.
Zero-shot neural scene segmentation, which reconstructs 3D neural segmentation field without manual annotations, serves as an effective way for scene understanding. However, existing models, especially the efficient 3D Gaussian-based methods, struggle to produce compact segmentation results. This issue stems primarily from their redundant learnable attributes assigned on individual Gaussians, leading to a lack of robustness against the 3D-inconsistencies in zero-shot generated raw labels. To address this problem, our work, named Compact Segmented 3D Gaussians (CoSegGaussians), proposes the Feature Unprojection and Fusion module as the segmentation field, which utilizes a shallow decoder generalizable for all Gaussians based on high-level features. Specifically, leveraging the learned Gaussian geometric parameters, semantic-aware image-based features are introduced into the scene via our unprojection technique. The lifted features, together with spatial information, are fed into the multi-scale aggregation decoder to generate segmentation identities for all Gaussians. Furthermore, we design CoSeg Loss to boost model robustness against 3D-inconsistent noises. Experimental results show that our model surpasses baselines on zero-shot semantic segmentation task, improving by ~10% mIoU over the best baseline. Code and more results will be available at https://David-Dou.github.io/CoSegGaussians.
In the conventional change detection (CD) pipeline, two manually registered and labeled remote sensing datasets serve as the input of the model for training and prediction. However, in realistic scenarios, data from different periods or sensors could fail to be aligned as a result of various coordinate systems. Geometric distortion caused by coordinate shifting remains a thorny issue for CD algorithms. In this paper, we propose a reusable self-supervised framework for bitemporal geometric distortion in CD tasks. The whole framework is composed of Pretext Representation Pre-training, Bitemporal Image Alignment, and Down-stream Decoder Fine-Tuning. With only single-stage pre-training, the key components of the framework can be reused for assistance in the bitemporal image alignment, while simultaneously enhancing the performance of the CD decoder. Experimental results in 2 large-scale realistic scenarios demonstrate that our proposed method can alleviate the bitemporal geometric distortion in CD tasks.
Reinforcement learning (RL) problems where the learner attempts to infer an unobserved reward from some feedback variables have been studied in several recent papers. The setting of Interaction-Grounded Learning (IGL) is an example of such feedback-based RL tasks where the learner optimizes the return by inferring latent binary rewards from the interaction with the environment. In the IGL setting, a relevant assumption used in the RL literature is that the feedback variable $Y$ is conditionally independent of the context-action $(X,A)$ given the latent reward $R$. In this work, we propose Variational Information-based IGL (VI-IGL) as an information-theoretic method to enforce the conditional independence assumption in the IGL-based RL problem. The VI-IGL framework learns a reward decoder using an information-based objective based on the conditional mutual information (MI) between $(X,A)$ and $Y$. To estimate and optimize the information-based terms for the continuous random variables in the RL problem, VI-IGL leverages the variational representation of mutual information to obtain a min-max optimization problem. Also, we extend the VI-IGL framework to general $f$-Information measures leading to the generalized $f$-VI-IGL framework for the IGL-based RL problems. We present numerical results on several reinforcement learning settings indicating an improved performance compared to the existing IGL-based RL algorithm.