We present ART$\boldsymbol{\cdot}$V, an efficient framework for auto-regressive video generation with diffusion models. Unlike existing methods that generate entire videos in one-shot, ART$\boldsymbol{\cdot}$V generates a single frame at a time, conditioned on the previous ones. The framework offers three distinct advantages. First, it only learns simple continual motions between adjacent frames, therefore avoiding modeling complex long-range motions that require huge training data. Second, it preserves the high-fidelity generation ability of the pre-trained image diffusion models by making only minimal network modifications. Third, it can generate arbitrarily long videos conditioned on a variety of prompts such as text, image or their combinations, making it highly versatile and flexible. To combat the common drifting issue in AR models, we propose masked diffusion model which implicitly learns which information can be drawn from reference images rather than network predictions, in order to reduce the risk of generating inconsistent appearances that cause drifting. Moreover, we further enhance generation coherence by conditioning it on the initial frame, which typically contains minimal noise. This is particularly useful for long video generation. When trained for only two weeks on four GPUs, ART$\boldsymbol{\cdot}$V already can generate videos with natural motions, rich details and a high level of aesthetic quality. Besides, it enables various appealing applications, e.g., composing a long video from multiple text prompts.
Deciphering the oracle bone script plays a significant role in Chinese archaeology and philology. However, it is significantly challenging due to the scarcity of oracle character images. To overcome this issue, we propose Diff-Oracle, based on diffusion models (DMs), to generate sufficient controllable oracle characters. In contrast to most DMs that rely on text prompts, we incorporate a style encoder to control style information during the generation process. This encoder extracts style prompts from existing oracle character images, where style details are converted from a CLIP model into a text embedding format. Inspired by ControlNet, we introduce a content encoder to capture desired content information from content images, ensuring the fidelity of character glyphs. To train Diff-Oracle effectively, we propose to obtain pixel-level paired oracle character images (i.e., style and content images) by a pre-trained image-to-image translation model. Extensive qualitative and quantitative experiments conducted on two benchmark datasets, Oracle-241 and OBC306, demonstrate that our Diff-Oracle outperforms existing generative methods in terms of image generation, further enhancing recognition accuracy. Source codes will be available.
In the evaluation of medical text generation, it is essential to scrutinize each piece of information and ensure the utmost accuracy of the evaluation. Existing evaluation metrics either focus on coarse-level evaluation that assigns one score for the whole generated output or rely on evaluation models trained on general domain, resulting in inaccuracies when adapted to the medical domain. To address these issues, we propose a set of factuality-centric evaluation aspects and design corresponding GPT-4-based metrics for medical text generation. We systematically compare these metrics with existing ones on clinical note generation and medical report summarization tasks, revealing low inter-metric correlation. A comprehensive human evaluation confirms that the proposed GPT-4-based metrics exhibit substantially higher agreement with human judgments than existing evaluation metrics. Our study contributes to the understanding of medical text generation evaluation and offers a more reliable alternative to existing metrics.
Speech emotions are crucial in human communication and are extensively used in fields like speech synthesis and natural language understanding. Most prior studies, such as speech emotion recognition, have categorized speech emotions into a fixed set of classes. Yet, emotions expressed in human speech are often complex, and categorizing them into predefined groups can be insufficient to adequately represent speech emotions. On the contrary, describing speech emotions directly by means of natural language may be a more effective approach. Regrettably, there are not many studies available that have focused on this direction. Therefore, this paper proposes a speech emotion captioning framework named SECap, aiming at effectively describing speech emotions using natural language. Owing to the impressive capabilities of large language models in language comprehension and text generation, SECap employs LLaMA as the text decoder to allow the production of coherent speech emotion captions. In addition, SECap leverages HuBERT as the audio encoder to extract general speech features and Q-Former as the Bridge-Net to provide LLaMA with emotion-related speech features. To accomplish this, Q-Former utilizes mutual information learning to disentangle emotion-related speech features and speech contents, while implementing contrastive learning to extract more emotion-related speech features. The results of objective and subjective evaluations demonstrate that: 1) the SECap framework outperforms the HTSAT-BART baseline in all objective evaluations; 2) SECap can generate high-quality speech emotion captions that attain performance on par with human annotators in subjective mean opinion score tests.
Equivariance is an important feature in machine learning, including language models. It ensures that any sequences of phrases with the same meanings are interpreted consistently. For example, the sentence 'There is a cat on the table' should be interpreted by language models as it is, regardless of variations in its token-level expression. Building on this insight, I propose a new theory suggesting that insufficient equivariance in language models can lead to hallucinations. According to this theory, which is both intuitive and novel, language models trained on relatively small datasets tend to misinterpret input texts and/or generate incorrect texts (i.e., hallucinations). To test this theory, I developed a toy model known as 'dancing men', which is a character-level substitution cipher. Additionally, I propose a novel technique based on the T5 (Text To Text Transfer Transformer) model to efficiently decipher these codes without relying on frequency analysis. I have found that this T5 model can almost completely solve the cipher, demonstrating its ability to acquire equivariance in this frame. This method could be scaled up to word-level and sentence-level substitution ciphers, analogous to large language models without tokenizers or dictionaries. This scalability makes it suitable for investigating the proposed link between inadequate equivariance acquisition and the emergence of hallucinations.
Prosody -- the suprasegmental component of speech, including pitch, loudness, and tempo -- carries critical aspects of meaning. However, the relationship between the information conveyed by prosody vs. by the words themselves remains poorly understood. We use large language models (LLMs) to estimate how much information is redundant between prosody and the words themselves. Using a large spoken corpus of English audiobooks, we extract prosodic features aligned to individual words and test how well they can be predicted from LLM embeddings, compared to non-contextual word embeddings. We find a high degree of redundancy between the information carried by the words and prosodic information across several prosodic features, including intensity, duration, pauses, and pitch contours. Furthermore, a word's prosodic information is redundant with both the word itself and the context preceding as well as following it. Still, we observe that prosodic features can not be fully predicted from text, suggesting that prosody carries information above and beyond the words. Along with this paper, we release a general-purpose data processing pipeline for quantifying the relationship between linguistic information and extra-linguistic features.
Multimodal sentiment analysis (MSA) leverages heterogeneous data sources to interpret the complex nature of human sentiments. Despite significant progress in multimodal architecture design, the field lacks comprehensive regularization methods. This paper introduces PowMix, a versatile embedding space regularizer that builds upon the strengths of unimodal mixing-based regularization approaches and introduces novel algorithmic components that are specifically tailored to multimodal tasks. PowMix is integrated before the fusion stage of multimodal architectures and facilitates intra-modal mixing, such as mixing text with text, to act as a regularizer. PowMix consists of five components: 1) a varying number of generated mixed examples, 2) mixing factor reweighting, 3) anisotropic mixing, 4) dynamic mixing, and 5) cross-modal label mixing. Extensive experimentation across benchmark MSA datasets and a broad spectrum of diverse architectural designs demonstrate the efficacy of PowMix, as evidenced by consistent performance improvements over baselines and existing mixing methods. An in-depth ablation study highlights the critical contribution of each PowMix component and how they synergistically enhance performance. Furthermore, algorithmic analysis demonstrates how PowMix behaves in different scenarios, particularly comparing early versus late fusion architectures. Notably, PowMix enhances overall performance without sacrificing model robustness or magnifying text dominance. It also retains its strong performance in situations of limited data. Our findings position PowMix as a promising versatile regularization strategy for MSA. Code will be made available.
The recent advancements in text-to-3D generation mark a significant milestone in generative models, unlocking new possibilities for creating imaginative 3D assets across various real-world scenarios. While recent advancements in text-to-3D generation have shown promise, they often fall short in rendering detailed and high-quality 3D models. This problem is especially prevalent as many methods base themselves on Score Distillation Sampling (SDS). This paper identifies a notable deficiency in SDS, that it brings inconsistent and low-quality updating direction for the 3D model, causing the over-smoothing effect. To address this, we propose a novel approach called Interval Score Matching (ISM). ISM employs deterministic diffusing trajectories and utilizes interval-based score matching to counteract over-smoothing. Furthermore, we incorporate 3D Gaussian Splatting into our text-to-3D generation pipeline. Extensive experiments show that our model largely outperforms the state-of-the-art in quality and training efficiency.
In the In-Context Learning (ICL) setup, various forms of label biases can manifest. One such manifestation is majority label bias, which arises when the distribution of labeled examples in the in-context samples is skewed towards one or more specific classes making Large Language Models (LLMs) more prone to predict those labels. Such discrepancies can arise from various factors, including logistical constraints, inherent biases in data collection methods, limited access to diverse data sources, etc. which are unavoidable in a real-world industry setup. In this work, we study the robustness of in-context learning in LLMs to shifts that occur due to majority label bias within the purview of text classification tasks. Prior works have shown that in-context learning with LLMs is susceptible to such biases. In our study, we go one level deeper and show that the robustness boundary varies widely for different models and tasks, with certain LLMs being highly robust (~90%) to majority label bias. Additionally, our findings also highlight the impact of model size and the richness of instructional prompts contributing towards model robustness. We restrict our study to only publicly available open-source models to ensure transparency and reproducibility.
In this work, we introduce Vision-Language Generative Pre-trained Transformer (VL-GPT), a transformer model proficient at concurrently perceiving and generating visual and linguistic data. VL-GPT achieves a unified pre-training approach for both image and text modalities by employing a straightforward auto-regressive objective, thereby enabling the model to process image and text as seamlessly as a language model processes text. To accomplish this, we initially propose a novel image tokenizer-detokenizer framework for visual data, specifically designed to transform raw images into a sequence of continuous embeddings and reconstruct them accordingly. In combination with the existing text tokenizer and detokenizer, this framework allows for the encoding of interleaved image-text data into a multimodal sequence, which can subsequently be fed into the transformer model. Consequently, VL-GPT can perform large-scale pre-training on multimodal corpora utilizing a unified auto-regressive objective (i.e., next-token prediction). Upon completion of pre-training, VL-GPT exhibits remarkable zero-shot and few-shot performance across a diverse range of vision and language understanding and generation tasks, including image captioning, visual question answering, text-to-image generation, and more. Additionally, the pre-trained model retrains in-context learning capabilities when provided with multimodal prompts. We further conduct instruction tuning on our VL-GPT, highlighting its exceptional potential for multimodal assistance. The source code and model weights shall be released.