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.
This work aims to provide an overview on the open-source multilanguage tool called StyloMetrix. It offers stylometric text representations that cover various aspects of grammar, syntax and lexicon. StyloMetrix covers four languages: Polish as the primary language, English, Ukrainian and Russian. The normalized output of each feature can become a fruitful course for machine learning models and a valuable addition to the embeddings layer for any deep learning algorithm. We strive to provide a concise, but exhaustive overview on the application of the StyloMetrix vectors as well as explain the sets of the developed linguistic features. The experiments have shown promising results in supervised content classification with simple algorithms as Random Forest Classifier, Voting Classifier, Logistic Regression and others. The deep learning assessments have unveiled the usefulness of the StyloMetrix vectors at enhancing an embedding layer extracted from Transformer architectures. The StyloMetrix has proven itself to be a formidable source for the machine learning and deep learning algorithms to execute different classification tasks.
The launch of ChatGPT at the end of 2022 generated large interest into possible applications of artificial intelligence in STEM education and among STEM professions. As a result many questions surrounding the capabilities of generative AI tools inside and outside of the classroom have been raised and are starting to be explored. This study examines the capabilities of ChatGPT within the discipline of mechanical engineering. It aims to examine use cases and pitfalls of such a technology in the classroom and professional settings. ChatGPT was presented with a set of questions from junior and senior level mechanical engineering exams provided at a large private university, as well as a set of practice questions for the Fundamentals of Engineering Exam (FE) in Mechanical Engineering. The responses of two ChatGPT models, one free to use and one paid subscription, were analyzed. The paper found that the subscription model (GPT-4) greatly outperformed the free version (GPT-3.5), achieving 76% correct vs 51% correct, but the limitation of text only input on both models makes neither likely to pass the FE exam. The results confirm findings in the literature with regards to types of errors and pitfalls made by ChatGPT. It was found that due to its inconsistency and a tendency to confidently produce incorrect answers the tool is best suited for users with expert knowledge.
The task of open-vocabulary object-centric image retrieval involves the retrieval of images containing a specified object of interest, delineated by an open-set text query. As working on large image datasets becomes standard, solving this task efficiently has gained significant practical importance. Applications include targeted performance analysis of retrieved images using ad-hoc queries and hard example mining during training. Recent advancements in contrastive-based open vocabulary systems have yielded remarkable breakthroughs, facilitating large-scale open vocabulary image retrieval. However, these approaches use a single global embedding per image, thereby constraining the system's ability to retrieve images containing relatively small object instances. Alternatively, incorporating local embeddings from detection pipelines faces scalability challenges, making it unsuitable for retrieval from large databases. In this work, we present a simple yet effective approach to object-centric open-vocabulary image retrieval. Our approach aggregates dense embeddings extracted from CLIP into a compact representation, essentially combining the scalability of image retrieval pipelines with the object identification capabilities of dense detection methods. We show the effectiveness of our scheme to the task by achieving significantly better results than global feature approaches on three datasets, increasing accuracy by up to 15 mAP points. We further integrate our scheme into a large scale retrieval framework and demonstrate our method's advantages in terms of scalability and interpretability.
Following the success of GPT4, there has been a surge in interest in multimodal large language model (MLLM) research. This line of research focuses on developing general-purpose LLMs through fine-tuning pre-trained LLMs and vision models. However, catastrophic forgetting, a notorious phenomenon where the fine-tuned model fails to retain similar performance compared to the pre-trained model, still remains an inherent problem in multimodal LLMs (MLLM). In this paper, we introduce EMT: Evaluating MulTimodality for evaluating the catastrophic forgetting in MLLMs, by treating each MLLM as an image classifier. We first apply EMT to evaluate several open-source fine-tuned MLLMs and we discover that almost all evaluated MLLMs fail to retain the same performance levels as their vision encoders on standard image classification tasks. Moreover, we continue fine-tuning LLaVA, an MLLM and utilize EMT to assess performance throughout the fine-tuning. Interestingly, our results suggest that early-stage fine-tuning on an image dataset improves performance across other image datasets, by enhancing the alignment of text and visual features. However, as fine-tuning proceeds, the MLLMs begin to hallucinate, resulting in a significant loss of generalizability, even when the image encoder remains frozen. Our results suggest that MLLMs have yet to demonstrate performance on par with their vision models on standard image classification tasks and the current MLLM fine-tuning procedure still has room for improvement.
This paper explores the potential of constructing an AI spoken dialogue system that "thinks how to respond" and "thinks how to speak" simultaneously, which more closely aligns with the human speech production process compared to the current cascade pipeline of independent chatbot and Text-to-Speech (TTS) modules. We hypothesize that Large Language Models (LLMs) with billions of parameters possess significant speech understanding capabilities and can jointly model dialogue responses and linguistic features. We conduct two sets of experiments: 1) Prosodic structure prediction, a typical front-end task in TTS, demonstrating the speech understanding ability of LLMs, and 2) Further integrating dialogue response and a wide array of linguistic features using a unified encoding format. Our results indicate that the LLM-based approach is a promising direction for building unified spoken dialogue systems.
Pretrained language models sometimes possess knowledge that we do not wish them to, including memorized personal information and knowledge that could be used to harm people. They can also output toxic or harmful text. To mitigate these safety and informational issues, we propose an attack-and-defense framework for studying the task of deleting sensitive information directly from model weights. We study direct edits to model weights because (1) this approach should guarantee that particular deleted information is never extracted by future prompt attacks, and (2) it should protect against whitebox attacks, which is necessary for making claims about safety/privacy in a setting where publicly available model weights could be used to elicit sensitive information. Our threat model assumes that an attack succeeds if the answer to a sensitive question is located among a set of B generated candidates, based on scenarios where the information would be insecure if the answer is among B candidates. Experimentally, we show that even state-of-the-art model editing methods such as ROME struggle to truly delete factual information from models like GPT-J, as our whitebox and blackbox attacks can recover "deleted" information from an edited model 38% of the time. These attacks leverage two key observations: (1) that traces of deleted information can be found in intermediate model hidden states, and (2) that applying an editing method for one question may not delete information across rephrased versions of the question. Finally, we provide new defense methods that protect against some extraction attacks, but we do not find a single universally effective defense method. Our results suggest that truly deleting sensitive information is a tractable but difficult problem, since even relatively low attack success rates have potentially severe societal implications for real-world deployment of language models.
Depression is a common mental disorder. Automatic depression detection tools using speech, enabled by machine learning, help early screening of depression. This paper addresses two limitations that may hinder the clinical implementations of such tools: noise resulting from segment-level labelling and a lack of model interpretability. We propose a bi-modal speech-level transformer to avoid segment-level labelling and introduce a hierarchical interpretation approach to provide both speech-level and sentence-level interpretations, based on gradient-weighted attention maps derived from all attention layers to track interactions between input features. We show that the proposed model outperforms a model that learns at a segment level ($p$=0.854, $r$=0.947, $F1$=0.947 compared to $p$=0.732, $r$=0.808, $F1$=0.768). For model interpretation, using one true positive sample, we show which sentences within a given speech are most relevant to depression detection; and which text tokens and Mel-spectrogram regions within these sentences are most relevant to depression detection. These interpretations allow clinicians to verify the validity of predictions made by depression detection tools, promoting their clinical implementations.
Large language models have made significant strides in natural language processing, paving the way for innovative applications including molecular representation and generation. However, most existing single-modality approaches cannot capture the abundant and complex information in molecular data. Here, we introduce GIT-Mol, a multi-modal large language model that integrates the structure Graph, Image, and Text information, including the Simplified Molecular Input Line Entry System (SMILES) and molecular captions. To facilitate the integration of multi-modal molecular data, we propose GIT-Former, a novel architecture capable of mapping all modalities into a unified latent space. Our study develops an innovative any-to-language molecular translation strategy and achieves a 10%-15% improvement in molecular captioning, a 5%-10% accuracy increase in property prediction, and a 20% boost in molecule generation validity compared to baseline or single-modality models.
Learning a recommender system model from an item's raw modality features (such as image, text, audio, etc.), called MoRec, has attracted growing interest recently. One key advantage of MoRec is that it can easily benefit from advances in other fields, such as natural language processing (NLP) and computer vision (CV). Moreover, it naturally supports transfer learning across different systems through modality features, known as transferable recommender systems, or TransRec. However, so far, TransRec has made little progress, compared to groundbreaking foundation models in the fields of NLP and CV. The lack of large-scale, high-quality recommendation datasets poses a major obstacle. To this end, we introduce NineRec, a TransRec dataset suite that includes a large-scale source domain recommendation dataset and nine diverse target domain recommendation datasets. Each item in NineRec is represented by a text description and a high-resolution cover image. With NineRec, we can implement TransRec models in an end-to-end training manner instead of using pre-extracted invariant features. We conduct a benchmark study and empirical analysis of TransRec using NineRec, and our findings provide several valuable insights. To support further research, we make our code, datasets, benchmarks, and leaderboards publicly available at https://github.com/anonymous?ninerec/NineRec.