Audio-Visual Speech-to-Speech Translation typically prioritizes improving translation quality and naturalness. However, an equally critical aspect in audio-visual content is lip-synchrony-ensuring that the movements of the lips match the spoken content-essential for maintaining realism in dubbed videos. Despite its importance, the inclusion of lip-synchrony constraints in AVS2S models has been largely overlooked. This study addresses this gap by integrating a lip-synchrony loss into the training process of AVS2S models. Our proposed method significantly enhances lip-synchrony in direct audio-visual speech-to-speech translation, achieving an average LSE-D score of 10.67, representing a 9.2% reduction in LSE-D over a strong baseline across four language pairs. Additionally, it maintains the naturalness and high quality of the translated speech when overlaid onto the original video, without any degradation in translation quality.

This week, we're looking back at our hellish 2024 and trying to figure out where to go from here. We began the year with enormous hype around artificial intelligence, but that's cooled off after seeing how useless many AI features have been. It's also clear that many companies, including Microsoft and Apple, are trying to push half-baked AI concepts onto users. Looking forward, we're expecting a rough few years for the tech industry (not to mention the world as a whole). Listen below or subscribe on your podcast app of choice. If you've got suggestions or topics you'd like covered on the show, be sure to email us or drop a note in the comments!

In federated learning, federated unlearning is a technique that provides clients with a rollback mechanism that allows them to withdraw their data contribution without training from scratch. However, existing research has not considered scenarios with skewed label distributions. Unfortunately, the unlearning of a client with skewed data usually results in biased models and makes it difficult to deliver high-quality service, complicating the recovery process. This paper proposes a recovery method of federated unlearning with skewed label distributions. Specifically, we first adopt a strategy that incorporates oversampling with deep learning to supplement the skewed class data for clients to perform recovery training, therefore enhancing the completeness of their local datasets. Afterward, a density-based denoising method is applied to remove noise from the generated data, further improving the quality of the remaining clients' datasets. Finally, all the remaining clients leverage the enhanced local datasets and engage in iterative training to effectively restore the performance of the unlearning model. Extensive evaluations on commonly used federated learning datasets with varying degrees of skewness show that our method outperforms baseline methods in restoring the performance of the unlearning model, particularly regarding accuracy on the skewed class.

Training data attribution (TDA) methods aim to attribute model outputs back to specific training examples, and the application of these methods to large language model (LLM) outputs could significantly advance model transparency and data curation. However, it has been challenging to date to apply these methods to the full scale of LLM pretraining. In this paper, we refine existing gradient-based methods to work effectively at scale, allowing us to retrieve influential examples for an 8B-parameter language model from a pretraining corpus of over 160B tokens with no need for subsampling or pre-filtering. Our method combines several techniques, including optimizer state correction, a task-specific Hessian approximation, and normalized encodings, which we find to be critical for performance at scale. In quantitative evaluations on a fact tracing task, our method performs best at identifying examples that influence model predictions, but classical, model-agnostic retrieval methods such as BM25 still perform better at finding passages which explicitly contain relevant facts. These results demonstrate a misalignment between factual *attribution* and causal *influence*. With increasing model size and training tokens, we find that influence more closely aligns with factual attribution. Finally, we examine different types of examples identified as influential by our method, finding that while many directly entail a particular fact, others support the same output by reinforcing priors on relation types, common entities, and names. We release our prompt set and model outputs, along with a web-based visualization tool to explore influential examples for factual predictions, commonsense reasoning, arithmetic, and open-ended generation for an 8B-parameter LLM.

The LLM-as-judge paradigm is increasingly being adopted for automated evaluation of model outputs. While LLM judges have shown promise on constrained evaluation tasks, closed source LLMs display critical shortcomings when deployed in real world applications due to challenges of fine grained metrics and explainability, while task specific evaluation models lack cross-domain generalization. We introduce GLIDER, a powerful 3B evaluator LLM that can score any text input and associated context on arbitrary user defined criteria. GLIDER shows higher Pearson's correlation than GPT-4o on FLASK and greatly outperforms prior evaluation models, achieving comparable performance to LLMs 17x its size. GLIDER supports fine-grained scoring, multilingual reasoning, span highlighting and was trained on 685 domains and 183 criteria. Extensive qualitative analysis shows that GLIDER scores are highly correlated with human judgments, with 91.3% human agreement. We have open-sourced GLIDER to facilitate future research.

We show that current open-source foundational LLMs possess instruction capability and German legal background knowledge that is sufficient for some legal analysis in an educational context. However, model capability breaks down in very specific tasks, such as the classification of "Gutachtenstil" appraisal style components, or with complex contexts, such as complete legal opinions. Even with extended context and effective prompting strategies, they cannot match the Bag-of-Words baseline. To combat this, we introduce a Retrieval Augmented Generation based prompt example selection method that substantially improves predictions in high data availability scenarios. We further evaluate the performance of pre-trained LLMs on two standard tasks for argument mining and automated essay scoring and find it to be more adequate. Throughout, pre-trained LLMs improve upon the baseline in scenarios with little or no labeled data with Chain-of-Thought prompting further helping in the zero-shot case.

Scaling the size of the memory for a 1.3 billion parameter base model (zero memory parameters corresponds to a dense model), trained to 1 trillion tokens. On the left, factual QA accuracy (exact match on NaturalQuestions and F1 score on TriviaQA), on the right task NLL (lower is better). Dashed lines show the performance of a 7B model trained on 2 trillion tokens with 10x more FLOPs. Pretrained language models encode vast amounts of information in their parameters (Roberts et al., 2020), and they can recall and use this information more accurately with increasing scale (Brown et al., 2020). For dense deep neural networks, which encode information primarily as weights of linear matrix transforms, this scaling of parameter size is directly coupled to an increase in computational and energy requirements. It is unclear if this is the most efficient solution to all information storage needs of language models. An important subset of information that language models need to learn are simple associations.

Existing pruning techniques for large language models (LLMs) targeting domain-specific applications typically follow a two-stage process: pruning the pretrained general-purpose LLMs and then fine-tuning the pruned LLMs on specific domains. However, the pruning decisions, derived from the pretrained weights, remain unchanged during fine-tuning, even if the weights have been updated. Therefore, such a combination of the pruning decisions and the finetuned weights may be suboptimal, leading to non-negligible performance degradation. To address these limitations, we propose ATP: All-in-One Tuning and Structural Pruning, a unified one-stage structural pruning and fine-tuning approach that dynamically identifies the current optimal substructure throughout the fine-tuning phase via a trainable pruning decision generator. Moreover, given the limited available data for domain-specific applications, Low-Rank Adaptation (LoRA) becomes a common technique to fine-tune the LLMs. In ATP, we introduce LoRA-aware forward and sparsity regularization to ensure that the substructures corresponding to the learned pruning decisions can be directly removed after the ATP process. ATP outperforms the state-of-the-art two-stage pruning methods on tasks in the legal and healthcare domains. More specifically, ATP recovers up to 88% and 91% performance of the dense model when pruning 40% parameters of LLaMA2-7B and LLaMA3-8B models, respectively.

Recent advancements in AI alignment techniques have significantly improved the alignment of large language models (LLMs) with static human preferences. However, the dynamic nature of human preferences can render some prior training data outdated or even erroneous, ultimately causing LLMs to deviate from contemporary human preferences and societal norms. Existing methodologies, whether they involve the curation of new data for continual alignment or the manual correction of outdated data for re-alignment, demand costly human resources. To address this challenge, we propose a novel approach, Large Language Model Behavior Correction with Influence Function Recall and Post-Training (LANCET), which requires no human involvement. LANCET consists of two phases: (1) using influence functions to identify the training data that significantly impact undesirable model outputs, and (2) applying an Influence function-driven Bregman Optimization (IBO) technique to adjust the model's behavior based on these influence distributions. Our experiments demonstrate that LANCET effectively and efficiently correct inappropriate behaviors of LLMs. Furthermore, LANCET can outperform methods that rely on collecting human preferences, and it enhances the interpretability of learning human preferences within LLMs.

When it comes to classifying child sexual abuse images, managing similar inter-class correlations and diverse intra-class correlations poses a significant challenge. Vision transformer models, unlike conventional deep convolutional network models, leverage a self-attention mechanism to capture global interactions among contextual local elements. This allows them to navigate through image patches effectively, avoiding incorrect correlations and reducing ambiguity in attention maps, thus proving their efficacy in computer vision tasks. Rather than directly analyzing child sexual abuse data, we constructed two datasets: one comprising clean and pornographic images and another with three classes, which additionally include images indicative of pornography, sourced from Reddit and Google Open Images data. In our experiments, we also employ an adult content image benchmark dataset. These datasets served as a basis for assessing the performance of vision transformer models in pornographic image classification. In our study, we conducted a comparative analysis between various popular vision transformer models and traditional pre-trained ResNet models. Furthermore, we compared them with established methods for sensitive image detection such as attention and metric learning based CNN and Bumble. The findings demonstrated that vision transformer networks surpassed the benchmark pre-trained models, showcasing their superior classification and detection capabilities in this task.