Large Vision-Language Models (LVLMs) are increasingly adept at generating contextually detailed and coherent responses from visual inputs. However, their application in multimodal decision-making and open-ended generation is hindered by a notable rate of hallucinations, where generated text inaccurately represents the visual contents. To address this issue, this paper introduces the Instruction Contrastive Decoding (ICD) method, a novel approach designed to reduce hallucinations during LVLM inference. Our method is inspired by our observation that what we call disturbance instructions significantly exacerbate hallucinations in multimodal fusion modules. ICD contrasts distributions from standard and instruction disturbance, thereby increasing alignment uncertainty and effectively subtracting hallucinated concepts from the original distribution. Through comprehensive experiments on discriminative benchmarks (POPE and MME) and a generative benchmark (LLaVa-Bench), we demonstrate that ICD significantly mitigates both object-level and attribute-level hallucinations. Moreover, our method not only addresses hallucinations but also significantly enhances the general perception and recognition capabilities of LVLMs.
Text-to-3D model adaptations have advanced static 3D model quality, but sequential 3D model generation, particularly for animatable objects with large motions, is still scarce. Our work proposes AnimatableDreamer, a text-to-4D generation framework capable of generating diverse categories of non-rigid objects while adhering to the object motions extracted from a monocular video. At its core, AnimatableDreamer is equipped with our novel optimization design dubbed Canonical Score Distillation (CSD), which simplifies the generation dimension from 4D to 3D by denoising over different frames in the time-varying camera spaces while conducting the distillation process in a unique canonical space shared per video. Concretely, CSD ensures that score gradients back-propagate to the canonical space through differentiable warping, hence guaranteeing the time-consistent generation and maintaining morphological plausibility across different poses. By lifting the 3D generator to 4D with warping functions, AnimatableDreamer offers a novel perspective on non-rigid 3D model generation and reconstruction. Besides, with inductive knowledge from a multi-view consistent diffusion model, CSD regularizes reconstruction from novel views, thus cyclically enhancing the generation process. Extensive experiments demonstrate the capability of our method in generating high-flexibility text-guided 3D models from the monocular video, while also showing improved reconstruction performance over typical non-rigid reconstruction methods. Project page https://AnimatableDreamer.github.io.
Large Language Models (LLMs) have emerged as dominant foundational models in modern NLP. However, the understanding of their prediction process and internal mechanisms, such as feed-forward networks and multi-head self-attention, remains largely unexplored. In this study, we probe LLMs from a human behavioral perspective, correlating values from LLMs with eye-tracking measures, which are widely recognized as meaningful indicators of reading patterns. Our findings reveal that LLMs exhibit a prediction pattern distinct from that of RNN-based LMs. Moreover, with the escalation of FFN layers, the capacity for memorization and linguistic knowledge encoding also surges until it peaks, subsequently pivoting to focus on comprehension capacity. The functions of self-attention are distributed across multiple heads. Lastly, we scrutinize the gate mechanisms, finding that they control the flow of information, with some gates promoting, while others eliminating information.
It is challenging to build a multi-singer high-fidelity singing voice synthesis system with cross-lingual ability by only using monolingual singers in the training stage. In this paper, we propose CrossSinger, which is a cross-lingual singing voice synthesizer based on Xiaoicesing2. Specifically, we utilize International Phonetic Alphabet to unify the representation for all languages of the training data. Moreover, we leverage conditional layer normalization to incorporate the language information into the model for better pronunciation when singers meet unseen languages. Additionally, gradient reversal layer (GRL) is utilized to remove singer biases included in lyrics since all singers are monolingual, which indicates singer's identity is implicitly associated with the text. The experiment is conducted on a combination of three singing voice datasets containing Japanese Kiritan dataset, English NUS-48E dataset, and one internal Chinese dataset. The result shows CrossSinger can synthesize high-fidelity songs for various singers with cross-lingual ability, including code-switch cases.
In the paper, we propose a novel approach for solving Bayesian inverse problems with physics-informed invertible neural networks (PI-INN). The architecture of PI-INN consists of two sub-networks: an invertible neural network (INN) and a neural basis network (NB-Net). The invertible map between the parametric input and the INN output with the aid of NB-Net is constructed to provide a tractable estimation of the posterior distribution, which enables efficient sampling and accurate density evaluation. Furthermore, the loss function of PI-INN includes two components: a residual-based physics-informed loss term and a new independence loss term. The presented independence loss term can Gaussianize the random latent variables and ensure statistical independence between two parts of INN output by effectively utilizing the estimated density function. Several numerical experiments are presented to demonstrate the efficiency and accuracy of the proposed PI-INN, including inverse kinematics, inverse problems of the 1-d and 2-d diffusion equations, and seismic traveltime tomography.
Many existing federated learning (FL) algorithms are designed for supervised learning tasks, assuming that the local data owned by the clients are well labeled. However, in many practical situations, it could be difficult and expensive to acquire complete data labels. Federated semi-supervised learning (Fed-SSL) is an attractive solution for fully utilizing both labeled and unlabeled data. Similar to that encountered in federated supervised learning, class distribution of labeled/unlabeled data could be non-i.i.d. among clients. Besides, in each client, the class distribution of labeled data may be distinct from that of unlabeled data. Unfortunately, both can severely jeopardize the FL performance. To address such challenging issues, we introduce two proper regularization terms that can effectively alleviate the class distribution mismatch problem in Fed-SSL. In addition, to overcome the non-i.i.d. data, we leverage the variance reduction and normalized averaging techniques to develop a novel Fed-SSL algorithm. Theoretically, we prove that the proposed method has a convergence rate of $\mathcal{O}(1/\sqrt{T})$, where $T$ is the number of communication rounds, even when the data distribution are non-i.i.d. among clients. To the best of our knowledge, it is the first formal convergence result for Fed-SSL problems. Numerical experiments based on MNIST data and CIFAR-10 data show that the proposed method can greatly improve the classification accuracy compared to baselines.
Person re-identification (Re-ID) aims to match the image frames which contain the same person in the surveillance videos. Most of the Re-ID algorithms conduct supervised training in some small labeled datasets, so directly deploying these trained models to the real-world large camera networks may lead to a poor performance due to underfitting. The significant difference between the source training dataset and the target testing dataset makes it challenging to incrementally optimize the model. To address this challenge, we propose a novel solution by transforming the unlabeled images in the target domain to fit the original classifier by using our proposed similarity preserved generative adversarial networks model, SimPGAN. Specifically, SimPGAN adopts the generative adversarial networks with the cycle consistency constraint to transform the unlabeled images in the target domain to the style of the source domain. Meanwhile, SimPGAN uses the similarity consistency loss, which is measured by a siamese deep convolutional neural network, to preserve the similarity of the transformed images of the same person. Comprehensive experiments based on multiple real surveillance datasets are conducted, and the results show that our algorithm is better than the state-of-the-art cross-dataset unsupervised person Re-ID algorithms.
Precise destination prediction of taxi trajectories can benefit many intelligent location based services such as accurate ad for passengers. Traditional prediction approaches, which treat trajectories as one-dimensional sequences and process them in single scale, fail to capture the diverse two-dimensional patterns of trajectories in different spatial scales. In this paper, we propose T-CONV which models trajectories as two-dimensional images, and adopts multi-layer convolutional neural networks to combine multi-scale trajectory patterns to achieve precise prediction. Furthermore, we conduct gradient analysis to visualize the multi-scale spatial patterns captured by T-CONV and extract the areas with distinct influence on the ultimate prediction. Finally, we integrate multiple local enhancement convolutional fields to explore these important areas deeply for better prediction. Comprehensive experiments based on real trajectory data show that T-CONV can achieve higher accuracy than the state-of-the-art methods.