Towards Achieving Human Parity on End-to-end Simultaneous Speech Translation via LLM Agent

In this paper, we present Cross Language Agent -- Simultaneous Interpretation, CLASI, a high-quality and human-like Simultaneous Speech Translation (SiST) System. Inspired by professional human interpreters, we utilize a novel data-driven read-write strategy to balance the translation quality and latency. To address the challenge of translating in-domain terminologies, CLASI employs a multi-modal retrieving module to obtain relevant information to augment the translation. Supported by LLMs, our approach can generate error-tolerated translation by considering the input audio, historical context, and retrieved information. Experimental results show that our system outperforms other systems by significant margins. Aligned with professional human interpreters, we evaluate CLASI with a better human evaluation metric, valid information proportion (VIP), which measures the amount of information that can be successfully conveyed to the listeners. In the real-world scenarios, where the speeches are often disfluent, informal, and unclear, CLASI achieves VIP of 81.3% and 78.0% for Chinese-to-English and English-to-Chinese translation directions, respectively. In contrast, state-of-the-art commercial or open-source systems only achieve 35.4% and 41.6%. On the extremely hard dataset, where other systems achieve under 13% VIP, CLASI can still achieve 70% VIP.

Core Knowledge Learning Framework for Graph Adaptation and Scalability Learning

Graph classification is a pivotal challenge in machine learning, especially within the realm of graph-based data, given its importance in numerous real-world applications such as social network analysis, recommendation systems, and bioinformatics. Despite its significance, graph classification faces several hurdles, including adapting to diverse prediction tasks, training across multiple target domains, and handling small-sample prediction scenarios. Current methods often tackle these challenges individually, leading to fragmented solutions that lack a holistic approach to the overarching problem. In this paper, we propose an algorithm aimed at addressing the aforementioned challenges. By incorporating insights from various types of tasks, our method aims to enhance adaptability, scalability, and generalizability in graph classification. Motivated by the recognition that the underlying subgraph plays a crucial role in GNN prediction, while the remainder is task-irrelevant, we introduce the Core Knowledge Learning (\method{}) framework for graph adaptation and scalability learning. \method{} comprises several key modules, including the core subgraph knowledge submodule, graph domain adaptation module, and few-shot learning module for downstream tasks. Each module is tailored to tackle specific challenges in graph classification, such as domain shift, label inconsistencies, and data scarcity. By learning the core subgraph of the entire graph, we focus on the most pertinent features for task relevance. Consequently, our method offers benefits such as improved model performance, increased domain adaptability, and enhanced robustness to domain variations. Experimental results demonstrate significant performance enhancements achieved by our method compared to state-of-the-art approaches.

EDDA: A Encoder-Decoder Data Augmentation Framework for Zero-Shot Stance Detection

Stance detection aims to determine the attitude expressed in text towards a given target. Zero-shot stance detection (ZSSD) has emerged to classify stances towards unseen targets during inference. Recent data augmentation techniques for ZSSD increase transferable knowledge between targets through text or target augmentation. However, these methods exhibit limitations. Target augmentation lacks logical connections between generated targets and source text, while text augmentation relies solely on training data, resulting in insufficient generalization. To address these issues, we propose an encoder-decoder data augmentation (EDDA) framework. The encoder leverages large language models and chain-of-thought prompting to summarize texts into target-specific if-then rationales, establishing logical relationships. The decoder generates new samples based on these expressions using a semantic correlation word replacement strategy to increase syntactic diversity. We also analyze the generated expressions to develop a rationale-enhanced network that fully utilizes the augmented data. Experiments on benchmark datasets demonstrate our approach substantially improves over state-of-the-art ZSSD techniques. The proposed EDDA framework increases semantic relevance and syntactic variety in augmented texts while enabling interpretable rationale-based learning.

Speech Translation with Large Language Models: An Industrial Practice

Given the great success of large language models (LLMs) across various tasks, in this paper, we introduce LLM-ST, a novel and effective speech translation model constructed upon a pre-trained LLM. By integrating the large language model (LLM) with a speech encoder and employing multi-task instruction tuning, LLM-ST can produce accurate timestamped transcriptions and translations, even from long audio inputs. Furthermore, our findings indicate that the implementation of Chain-of-Thought (CoT) prompting can yield advantages in the context of LLM-ST. Through rigorous experimentation on English and Chinese datasets, we showcase the exceptional performance of LLM-ST, establishing a new benchmark in the field of speech translation. Demo: https://speechtranslation.github.io/llm-st/.

RepCodec: A Speech Representation Codec for Speech Tokenization

With recent rapid growth of large language models (LLMs), discrete speech tokenization has played an important role for injecting speech into LLMs. However, this discretization gives rise to a loss of information, consequently impairing overall performance. To improve the performance of these discrete speech tokens, we present RepCodec, a novel speech representation codec for semantic speech tokenization. In contrast to audio codecs which reconstruct the raw audio, RepCodec learns a vector quantization codebook through reconstructing speech representations from speech encoders like HuBERT or data2vec. Together, the speech encoder, the codec encoder and the vector quantization codebook form a pipeline for converting speech waveforms into semantic tokens. The extensive experiments illustrate that RepCodec, by virtue of its enhanced information retention capacity, significantly outperforms the widely used k-means clustering approach in both speech understanding and generation. Furthermore, this superiority extends across various speech encoders and languages, affirming the robustness of RepCodec. We believe our method can facilitate large language modeling research on speech processing.

Fast Adversarial Training with Adaptive Step Size

While adversarial training and its variants have shown to be the most effective algorithms to defend against adversarial attacks, their extremely slow training process makes it hard to scale to large datasets like ImageNet. The key idea of recent works to accelerate adversarial training is to substitute multi-step attacks (e.g., PGD) with single-step attacks (e.g., FGSM). However, these single-step methods suffer from catastrophic overfitting, where the accuracy against PGD attack suddenly drops to nearly 0% during training, destroying the robustness of the networks. In this work, we study the phenomenon from the perspective of training instances. We show that catastrophic overfitting is instance-dependent and fitting instances with larger gradient norm is more likely to cause catastrophic overfitting. Based on our findings, we propose a simple but effective method, Adversarial Training with Adaptive Step size (ATAS). ATAS learns an instancewise adaptive step size that is inversely proportional to its gradient norm. The theoretical analysis shows that ATAS converges faster than the commonly adopted non-adaptive counterparts. Empirically, ATAS consistently mitigates catastrophic overfitting and achieves higher robust accuracy on CIFAR10, CIFAR100 and ImageNet when evaluated on various adversarial budgets.

Black-box Prompt Learning for Pre-trained Language Models

Domain-specific fine-tuning strategies for large pre-trained models received vast attention in recent years. In previously studied settings, the model architectures and parameters are tunable or at least visible, which we refer to as white-box settings. This work considers a new scenario, where we do not have access to a pre-trained model, except for its outputs given inputs, and we call this problem black-box fine-tuning. To illustrate our approach, we first introduce the black-box setting formally on text classification, where the pre-trained model is not only frozen but also invisible. We then propose our solution black-box prompt, a new technique in the prompt-learning family, which can leverage the knowledge learned by pre-trained models from the pre-training corpus. Our experiments demonstrate that the proposed method achieved the state-of-the-art performance on eight datasets. Further analyses on different human-designed objectives, prompt lengths, and intuitive explanations demonstrate the robustness and flexibility of our method.

On the Impact of Hard Adversarial Instances on Overfitting in Adversarial Training

Adversarial training is a popular method to robustify models against adversarial attacks. However, it exhibits much more severe overfitting than training on clean inputs. In this work, we investigate this phenomenon from the perspective of training instances, i.e., training input-target pairs. Based on a quantitative metric measuring instances' difficulty, we analyze the model's behavior on training instances of different difficulty levels. This lets us show that the decay in generalization performance of adversarial training is a result of the model's attempt to fit hard adversarial instances. We theoretically verify our observations for both linear and general nonlinear models, proving that models trained on hard instances have worse generalization performance than ones trained on easy instances. Furthermore, we prove that the difference in the generalization gap between models trained by instances of different difficulty levels increases with the size of the adversarial budget. Finally, we conduct case studies on methods mitigating adversarial overfitting in several scenarios. Our analysis shows that methods successfully mitigating adversarial overfitting all avoid fitting hard adversarial instances, while ones fitting hard adversarial instances do not achieve true robustness.

Few-Shot Human Motion Transfer by Personalized Geometry and Texture Modeling

We present a new method for few-shot human motion transfer that achieves realistic human image generation with only a small number of appearance inputs. Despite recent advances in single person motion transfer, prior methods often require a large number of training images and take long training time. One promising direction is to perform few-shot human motion transfer, which only needs a few of source images for appearance transfer. However, it is particularly challenging to obtain satisfactory transfer results. In this paper, we address this issue by rendering a human texture map to a surface geometry (represented as a UV map), which is personalized to the source person. Our geometry generator combines the shape information from source images, and the pose information from 2D keypoints to synthesize the personalized UV map. A texture generator then generates the texture map conditioned on the texture of source images to fill out invisible parts. Furthermore, we may fine-tune the texture map on the manifold of the texture generator from a few source images at the test time, which improves the quality of the texture map without over-fitting or artifacts. Extensive experiments show the proposed method outperforms state-of-the-art methods both qualitatively and quantitatively. Our code is available at https://github.com/HuangZhiChao95/FewShotMotionTransfer.

CorrAttack: Black-box Adversarial Attack with Structured Search

We present a new method for score-based adversarial attack, where the attacker queries the loss-oracle of the target model. Our method employs a parameterized search space with a structure that captures the relationship of the gradient of the loss function. We show that searching over the structured space can be approximated by a time-varying contextual bandits problem, where the attacker takes feature of the associated arm to make modifications of the input, and receives an immediate reward as the reduction of the loss function. The time-varying contextual bandits problem can then be solved by a Bayesian optimization procedure, which can take advantage of the features of the structured action space. The experiments on ImageNet and the Google Cloud Vision API demonstrate that the proposed method achieves the state of the art success rates and query efficiencies for both undefended and defended models.