The scene text removal (STR) task aims to remove text regions and recover the background smoothly in images for private information protection. Most existing STR methods adopt encoder-decoder-based CNNs, with direct copies of the features in the skip connections. However, the encoded features contain both text texture and structure information. The insufficient utilization of text features hampers the performance of background reconstruction in text removal regions. To tackle these problems, we propose a novel Feature Erasing and Transferring (FET) mechanism to reconfigure the encoded features for STR in this paper. In FET, a Feature Erasing Module (FEM) is designed to erase text features. An attention module is responsible for generating the feature similarity guidance. The Feature Transferring Module (FTM) is introduced to transfer the corresponding features in different layers based on the attention guidance. With this mechanism, a one-stage, end-to-end trainable network called FETNet is constructed for scene text removal. In addition, to facilitate research on both scene text removal and segmentation tasks, we introduce a novel dataset, Flickr-ST, with multi-category annotations. A sufficient number of experiments and ablation studies are conducted on the public datasets and Flickr-ST. Our proposed method achieves state-of-the-art performance using most metrics, with remarkably higher quality scene text removal results. The source code of our work is available at: \href{https://github.com/GuangtaoLyu/FETNet}{https://github.com/GuangtaoLyu/FETNet.
As Large Language Models are deployed within Artificial Intelligence systems, that are increasingly integrated with human society, it becomes more important than ever to study their internal structures. Higher level abilities of LLMs such as GPT-3.5 emerge in large part due to informative language representations they induce from raw text data during pre-training on trillions of words. These embeddings exist in vector spaces of several thousand dimensions, and their processing involves mapping between multiple vector spaces, with total number of parameters on the order of trillions. Furthermore, these language representations are induced by gradient optimization, resulting in a black box system that is hard to interpret. In this paper, we take a look at the topological structure of neuronal activity in the "brain" of Chat-GPT's foundation language model, and analyze it with respect to a metric representing the notion of fairness. We develop a novel approach to visualize GPT's moral dimensions. We first compute a fairness metric, inspired by social psychology literature, to identify factors that typically influence fairness assessments in humans, such as legitimacy, need, and responsibility. Subsequently, we summarize the manifold's shape using a lower-dimensional simplicial complex, whose topology is derived from this metric. We color it with a heat map associated with this fairness metric, producing human-readable visualizations of the high-dimensional sentence manifold. Our results show that sentence embeddings based on GPT-3.5 can be decomposed into two submanifolds corresponding to fair and unfair moral judgments. This indicates that GPT-based language models develop a moral dimension within their representation spaces and induce an understanding of fairness during their training process.
Text correction, especially the semantic correction of more widely used scenes, is strongly required to improve, for the fluency and writing efficiency of the text. An adversarial multi-task learning method is proposed to enhance the modeling and detection ability of character polysemy in Chinese sentence context. Wherein, two models, the masked language model and scoring language model, are introduced as a pair of not only coupled but also adversarial learning tasks. Moreover, the Monte Carlo tree search strategy and a policy network are introduced to accomplish the efficient Chinese text correction task with semantic detection. The experiments are executed on three datasets and five comparable methods, and the experimental results show that our method can obtain good performance in Chinese text correction task for better semantic rationality.
Text-to-3D modelling has seen exciting progress by combining generative text-to-image models with image-to-3D methods like Neural Radiance Fields. DreamFusion recently achieved high-quality results but requires a lengthy, per-prompt optimization to create 3D objects. To address this, we amortize optimization over text prompts by training on many prompts simultaneously with a unified model, instead of separately. With this, we share computation across a prompt set, training in less time than per-prompt optimization. Our framework - Amortized text-to-3D (ATT3D) - enables knowledge-sharing between prompts to generalize to unseen setups and smooth interpolations between text for novel assets and simple animations.
This report presents MagicAvatar, a framework for multimodal video generation and animation of human avatars. Unlike most existing methods that generate avatar-centric videos directly from multimodal inputs (e.g., text prompts), MagicAvatar explicitly disentangles avatar video generation into two stages: (1) multimodal-to-motion and (2) motion-to-video generation. The first stage translates the multimodal inputs into motion/ control signals (e.g., human pose, depth, DensePose); while the second stage generates avatar-centric video guided by these motion signals. Additionally, MagicAvatar supports avatar animation by simply providing a few images of the target person. This capability enables the animation of the provided human identity according to the specific motion derived from the first stage. We demonstrate the flexibility of MagicAvatar through various applications, including text-guided and video-guided avatar generation, as well as multimodal avatar animation.
Multimodal learning often outperforms its unimodal counterparts by exploiting unimodal contributions and cross-modal interactions. However, focusing only on integrating multimodal features into a unified comprehensive representation overlooks the unimodal characteristics. In real data, the contributions of modalities can vary from instance to instance, and they often reinforce or conflict with each other. In this study, we introduce a novel \text{MultiModal} loss paradigm for multimodal learning, which subgroups instances according to their unimodal contributions. \text{MultiModal} loss can prevent inefficient learning caused by overfitting and efficiently optimize multimodal models. On synthetic data, \text{MultiModal} loss demonstrates improved classification performance by subgrouping difficult instances within certain modalities. On four real multimodal datasets, our loss is empirically shown to improve the performance of recent models. Ablation studies verify the effectiveness of our loss. Additionally, we show that our loss generates a reliable prediction score for each modality, which is essential for subgrouping. Our \text{MultiModal} loss is a novel loss function to subgroup instances according to the contribution of modalities in multimodal learning and is applicable to a variety of multimodal models with unimodal decisions. Our code is available at https://github.com/SehwanMoon/MultiModalLoss.
Multimodal stock trading volume movement prediction with stock-related news is one of the fundamental problems in the financial area. Existing multimodal works that train models from scratch face the problem of lacking universal knowledge when modeling financial news. In addition, the models ability may be limited by the lack of domain-related knowledge due to insufficient data in the datasets. To handle this issue, we propose the Prompt-based MUltimodal Stock volumE prediction model (ProMUSE) to process text and time series modalities. We use pre-trained language models for better comprehension of financial news and adopt prompt learning methods to leverage their capability in universal knowledge to model textual information. Besides, simply fusing two modalities can cause harm to the unimodal representations. Thus, we propose a novel cross-modality contrastive alignment while reserving the unimodal heads beside the fusion head to mitigate this problem. Extensive experiments demonstrate that our proposed ProMUSE outperforms existing baselines. Comprehensive analyses further validate the effectiveness of our architecture compared to potential variants and learning mechanisms.
When training powerful AI systems to perform complex tasks, it may be challenging to provide training signals which are robust to optimization. One concern is measurement tampering, where the AI system manipulates multiple measurements to create the illusion of good results instead of achieving the desired outcome. In this work, we build four new text-based datasets to evaluate measurement tampering detection techniques on large language models. Concretely, given sets of text inputs and measurements aimed at determining if some outcome occurred, as well as a base model able to accurately predict measurements, the goal is to determine if examples where all measurements indicate the outcome actually had the outcome occur, or if this was caused by measurement tampering. We demonstrate techniques that outperform simple baselines on most datasets, but don't achieve maximum performance. We believe there is significant room for improvement for both techniques and datasets, and we are excited for future work tackling measurement tampering.
Diffusion models have emerged as a powerful paradigm for generation, obtaining strong performance in various domains with continuous-valued inputs. Despite the promises of fully non-autoregressive text generation, applying diffusion models to natural language remains challenging due to its discrete nature. In this work, we propose Text-to-text Self-conditioned Simplex Diffusion (TESS), a text diffusion model that is fully non-autoregressive, employs a new form of self-conditioning, and applies the diffusion process on the logit simplex space rather than the typical learned embedding space. Through extensive experiments on natural language understanding and generation tasks including summarization, text simplification, paraphrase generation, and question generation, we demonstrate that TESS outperforms state-of-the-art non-autoregressive models and is competitive with pretrained autoregressive sequence-to-sequence models.
In this paper, we present MuLanTTS, the Microsoft end-to-end neural text-to-speech (TTS) system designed for the Blizzard Challenge 2023. About 50 hours of audiobook corpus for French TTS as hub task and another 2 hours of speaker adaptation as spoke task are released to build synthesized voices for different test purposes including sentences, paragraphs, homographs, lists, etc. Building upon DelightfulTTS, we adopt contextual and emotion encoders to adapt the audiobook data to enrich beyond sentences for long-form prosody and dialogue expressiveness. Regarding the recording quality, we also apply denoise algorithms and long audio processing for both corpora. For the hub task, only the 50-hour single speaker data is used for building the TTS system, while for the spoke task, a multi-speaker source model is used for target speaker fine tuning. MuLanTTS achieves mean scores of quality assessment 4.3 and 4.5 in the respective tasks, statistically comparable with natural speech while keeping good similarity according to similarity assessment. The excellent and similarity in this year's new and dense statistical evaluation show the effectiveness of our proposed system in both tasks.