Recent years have witnessed a significant increase in the performance of Vision and Language tasks. Foundational Vision-Language Models (VLMs), such as CLIP, have been leveraged in multiple settings and demonstrated remarkable performance across several tasks. Such models excel at object-centric recognition yet learn text representations that seem invariant to word order, failing to compose known concepts in novel ways. However, no evidence exists that any VLM, including large-scale single-stream models such as GPT-4V, identifies compositions successfully. In this paper, we introduce a framework to significantly improve the ability of existing models to encode compositional language, with over 10% absolute improvement on compositionality benchmarks, while maintaining or improving the performance on standard object-recognition and retrieval benchmarks. Our code and pre-trained models are publicly available at https://github.com/netflix/clove.
Text matching systems have become a fundamental service in most searching platforms. For instance, they are responsible for matching user queries to relevant candidate items, or rewriting the user-input query to a pre-selected high-performing one for a better search experience. In practice, both the queries and items often contain multiple attributes, such as the category of the item and the location mentioned in the query, which represent condensed key information that is helpful for matching. However, most of the existing works downplay the effectiveness of attributes by integrating them into text representations as supplementary information. Hence, in this work, we focus on exploring the relationship between the attributes from two sides. Since attributes from two ends are often not aligned in terms of number and type, we propose to exploit the benefit of attributes by multiple-intent modeling. The intents extracted from attributes summarize the diverse needs of queries and provide rich content of items, which are more refined and abstract, and can be aligned for paired inputs. Concretely, we propose a multi-intent attribute-aware matching model (MIM), which consists of three main components: attribute-aware encoder, multi-intent modeling, and intent-aware matching. In the attribute-aware encoder, the text and attributes are weighted and processed through a scaled attention mechanism with regard to the attributes' importance. Afterward, the multi-intent modeling extracts intents from two ends and aligns them. Herein, we come up with a distribution loss to ensure the learned intents are diverse but concentrated, and a kullback-leibler divergence loss that aligns the learned intents. Finally, in the intent-aware matching, the intents are evaluated by a self-supervised masking task, and then incorporated to output the final matching result.
While (large) language models have significantly improved over the last years, they still struggle to sensibly process long sequences found, e.g., in books, due to the quadratic scaling of the underlying attention mechanism. To address this, we propose NextLevelBERT, a Masked Language Model operating not on tokens, but on higher-level semantic representations in the form of text embeddings. We pretrain NextLevelBERT to predict the vector representation of entire masked text chunks and evaluate the effectiveness of the resulting document vectors on three task types: 1) Semantic Textual Similarity via zero-shot document embeddings, 2) Long document classification, 3) Multiple-choice question answering. We find that next level Masked Language Modeling is an effective technique to tackle long-document use cases and can outperform much larger embedding models as long as the required level of detail is not too high. We make model and code available.
Word order is an important concept in natural language, and in this work, we study how word order affects the induction of world knowledge from raw text using language models. We use word analogies to probe for such knowledge. Specifically, in addition to the natural word order, we first respectively extract texts of six fixed word orders from five languages and then pretrain the language models on these texts. Finally, we analyze the experimental results of the fixed word orders on word analogies and show that i) certain fixed word orders consistently outperform or underperform others, though the specifics vary across languages, and ii) the Wov2Lex hypothesis is not hold in pre-trained language models, and the natural word order typically yields mediocre results. The source code will be made publicly available at https://github.com/lshowway/probing_by_analogy.
Editing signals using large pre-trained models, in a zero-shot manner, has recently seen rapid advancements in the image domain. However, this wave has yet to reach the audio domain. In this paper, we explore two zero-shot editing techniques for audio signals, which use DDPM inversion on pre-trained diffusion models. The first, adopted from the image domain, allows text-based editing. The second, is a novel approach for discovering semantically meaningful editing directions without supervision. When applied to music signals, this method exposes a range of musically interesting modifications, from controlling the participation of specific instruments to improvisations on the melody. Samples and code can be found on our examples page in https://hilamanor.github.io/AudioEditing/ .
This paper introduces SceneCraft, a Large Language Model (LLM) Agent converting text descriptions into Blender-executable Python scripts which render complex scenes with up to a hundred 3D assets. This process requires complex spatial planning and arrangement. We tackle these challenges through a combination of advanced abstraction, strategic planning, and library learning. SceneCraft first models a scene graph as a blueprint, detailing the spatial relationships among assets in the scene. SceneCraft then writes Python scripts based on this graph, translating relationships into numerical constraints for asset layout. Next, SceneCraft leverages the perceptual strengths of vision-language foundation models like GPT-V to analyze rendered images and iteratively refine the scene. On top of this process, SceneCraft features a library learning mechanism that compiles common script functions into a reusable library, facilitating continuous self-improvement without expensive LLM parameter tuning. Our evaluation demonstrates that SceneCraft surpasses existing LLM-based agents in rendering complex scenes, as shown by its adherence to constraints and favorable human assessments. We also showcase the broader application potential of SceneCraft by reconstructing detailed 3D scenes from the Sintel movie and guiding a video generative model with generated scenes as intermediary control signal.
Retrieval-Augmented Generation (RAG) improves pre-trained models by incorporating external knowledge at test time to enable customized adaptation. We study the risk of datastore leakage in Retrieval-In-Context RAG Language Models (LMs). We show that an adversary can exploit LMs' instruction-following capabilities to easily extract text data verbatim from the datastore of RAG systems built with instruction-tuned LMs via prompt injection. The vulnerability exists for a wide range of modern LMs that span Llama2, Mistral/Mixtral, Vicuna, SOLAR, WizardLM, Qwen1.5, and Platypus2, and the exploitability exacerbates as the model size scales up. Extending our study to production RAG models GPTs, we design an attack that can cause datastore leakage with a 100% success rate on 25 randomly selected customized GPTs with at most 2 queries, and we extract text data verbatim at a rate of 41% from a book of 77,000 words and 3% from a corpus of 1,569,000 words by prompting the GPTs with only 100 queries generated by themselves.
Image generation models are poised to become ubiquitous in a range of applications. These models are often fine-tuned and evaluated using human quality judgments that assume a universal standard, failing to consider the subjectivity of such tasks. To investigate how to quantify subjectivity, and the scale of its impact, we measure how assessments differ among human annotators across different use cases. Simulating the effects of ordinarily latent elements of annotators subjectivity, we contrive a set of motivations (t-shirt graphics, presentation visuals, and phone background images) to contextualize a set of crowdsourcing tasks. Our results show that human evaluations of images vary within individual contexts and across combinations of contexts. Three key factors affecting this subjectivity are image appearance, image alignment with text, and representation of objects mentioned in the text. Our study highlights the importance of taking individual users and contexts into account, both when building and evaluating generative models
Likelihood-based deep generative models such as score-based diffusion models and variational autoencoders are state-of-the-art machine learning models approximating high-dimensional distributions of data such as images, text, or audio. One of many downstream tasks they can be naturally applied to is out-of-distribution (OOD) detection. However, seminal work by Nalisnick et al. which we reproduce showed that deep generative models consistently infer higher log-likelihoods for OOD data than data they were trained on, marking an open problem. In this work, we analyse using the gradient of a data point with respect to the parameters of the deep generative model for OOD detection, based on the simple intuition that OOD data should have larger gradient norms than training data. We formalise measuring the size of the gradient as approximating the Fisher information metric. We show that the Fisher information matrix (FIM) has large absolute diagonal values, motivating the use of chi-square distributed, layer-wise gradient norms as features. We combine these features to make a simple, model-agnostic and hyperparameter-free method for OOD detection which estimates the joint density of the layer-wise gradient norms for a given data point. We find that these layer-wise gradient norms are weakly correlated, rendering their combined usage informative, and prove that the layer-wise gradient norms satisfy the principle of (data representation) invariance. Our empirical results indicate that this method outperforms the Typicality test for most deep generative models and image dataset pairings.
We introduce a text-to-speech (TTS) model called BASE TTS, which stands for $\textbf{B}$ig $\textbf{A}$daptive $\textbf{S}$treamable TTS with $\textbf{E}$mergent abilities. BASE TTS is the largest TTS model to-date, trained on 100K hours of public domain speech data, achieving a new state-of-the-art in speech naturalness. It deploys a 1-billion-parameter autoregressive Transformer that converts raw texts into discrete codes ("speechcodes") followed by a convolution-based decoder which converts these speechcodes into waveforms in an incremental, streamable manner. Further, our speechcodes are built using a novel speech tokenization technique that features speaker ID disentanglement and compression with byte-pair encoding. Echoing the widely-reported "emergent abilities" of large language models when trained on increasing volume of data, we show that BASE TTS variants built with 10K+ hours and 500M+ parameters begin to demonstrate natural prosody on textually complex sentences. We design and share a specialized dataset to measure these emergent abilities for text-to-speech. We showcase state-of-the-art naturalness of BASE TTS by evaluating against baselines that include publicly available large-scale text-to-speech systems: YourTTS, Bark and TortoiseTTS. Audio samples generated by the model can be heard at https://amazon-ltts-paper.com/.