The field of Sign Language Production (SLP) lacked a large-scale, pre-trained model based on deep learning for continuous American Sign Language (ASL) production in the past decade. This limitation hampers communication for all individuals with disabilities relying on ASL. To address this issue, we undertook the secondary development and utilization of How2Sign, one of the largest publicly available ASL datasets. Despite its significance, prior researchers in the field of sign language have not effectively employed this corpus due to the intricacies involved in American Sign Language Production (ASLP). To conduct large-scale ASLP, we propose SignDiff based on the latest work in related fields, which is a dual-condition diffusion pre-training model that can generate human sign language speakers from a skeleton pose. SignDiff has a novel Frame Reinforcement Network called FR-Net, similar to dense human pose estimation work, which enhances the correspondence between text lexical symbols and sign language dense pose frames reduce the occurrence of multiple fingers in the diffusion model. In addition, our ASLP method proposes two new improved modules and a new loss function to improve the accuracy and quality of sign language skeletal posture and enhance the ability of the model to train on large-scale data. We propose the first baseline for ASL production and report the scores of 17.19 and 12.85 on BLEU-4 on the How2Sign dev/test sets. We also evaluated our model on the previous mainstream dataset called PHOENIX14T, and the main experiments achieved the results of SOTA. In addition, our image quality far exceeds all previous results by 10 percentage points on the SSIM indicator. Finally, we conducted ablation studies and qualitative evaluations for discussion.
Despite their promising performance across various natural language processing (NLP) tasks, current NLP systems are vulnerable to textual adversarial attacks. To defend against these attacks, most existing methods apply adversarial training by incorporating adversarial examples. However, these methods have to rely on ground-truth labels to generate adversarial examples, rendering it impractical for large-scale model pre-training which is commonly used nowadays for NLP and many other tasks. In this paper, we propose a novel learning framework called SCAT (Self-supervised Contrastive Learning via Adversarial Training), which can learn robust representations without requiring labeled data. Specifically, SCAT modifies random augmentations of the data in a fully labelfree manner to generate adversarial examples. Adversarial training is achieved by minimizing the contrastive loss between the augmentations and their adversarial counterparts. We evaluate SCAT on two text classification datasets using two state-of-the-art attack schemes proposed recently. Our results show that SCAT can not only train robust language models from scratch, but it can also significantly improve the robustness of existing pre-trained language models. Moreover, to demonstrate its flexibility, we show that SCAT can also be combined with supervised adversarial training to further enhance model robustness.
We introduce Ghostbuster, a state-of-the-art system for detecting AI-generated text. Our method works by passing documents through a series of weaker language models and running a structured search over possible combinations of their features, then training a classifier on the selected features to determine if the target document was AI-generated. Crucially, Ghostbuster does not require access to token probabilities from the target model, making it useful for detecting text generated by black-box models or unknown model versions. In conjunction with our model, we release three new datasets of human and AI-generated text as detection benchmarks that cover multiple domains (student essays, creative fiction, and news) and task setups: document-level detection, author identification, and a challenge task of paragraph-level detection. Ghostbuster averages 99.1 F1 across all three datasets on document-level detection, outperforming previous approaches such as GPTZero and DetectGPT by up to 32.7 F1.
Recently, visual-language learning has shown great potential in enhancing visual-based person re-identification (ReID). Existing visual-language learning-based ReID methods often focus on whole-body scale image-text feature alignment, while neglecting supervisions on fine-grained part features. This choice simplifies the learning process but cannot guarantee within-part feature semantic consistency thus hindering the final performance. Therefore, we propose to enhance fine-grained visual features with part-informed language supervision for ReID tasks. The proposed method, named Part-Informed Visual-language Learning ($\pi$-VL), suggests that (i) a human parsing-guided prompt tuning strategy and (ii) a hierarchical fusion-based visual-language alignment paradigm play essential roles in ensuring within-part feature semantic consistency. Specifically, we combine both identity labels and parsing maps to constitute pixel-level text prompts and fuse multi-stage visual features with a light-weight auxiliary head to perform fine-grained image-text alignment. As a plug-and-play and inference-free solution, our $\pi$-VL achieves substantial improvements over previous state-of-the-arts on four common-used ReID benchmarks, especially reporting 90.3% Rank-1 and 76.5% mAP for the most challenging MSMT17 database without bells and whistles.
The recent progress in diffusion-based text-to-image generation models has significantly expanded generative capabilities via conditioning the text descriptions. However, since relying solely on text prompts is still restrictive for fine-grained customization, we aim to extend the boundaries of conditional generation to incorporate diverse types of modalities, e.g., sketch, box, and style embedding, simultaneously. We thus design a multimodal text-to-image diffusion model, coined as DiffBlender, that achieves the aforementioned goal in a single model by training only a few small hypernetworks. DiffBlender facilitates a convenient scaling of input modalities, without altering the parameters of an existing large-scale generative model to retain its well-established knowledge. Furthermore, our study sets new standards for multimodal generation by conducting quantitative and qualitative comparisons with existing approaches. By diversifying the channels of conditioning modalities, DiffBlender faithfully reflects the provided information or, in its absence, creates imaginative generation.
The legality of training language models (LMs) on copyrighted or otherwise restricted data is under intense debate. However, as we show, model performance significantly degrades if trained only on low-risk text (e.g., out-of-copyright books or government documents), due to its limited size and domain coverage. We present SILO, a new language model that manages this risk-performance tradeoff during inference. SILO is built by (1) training a parametric LM on Open License Corpus (OLC), a new corpus we curate with 228B tokens of public domain and permissively licensed text and (2) augmenting it with a more general and easily modifiable nonparametric datastore (e.g., containing copyrighted books or news) that is only queried during inference. The datastore allows use of high-risk data without training on it, supports sentence-level data attribution, and enables data producers to opt out from the model by removing content from the store. These capabilities can foster compliance with data-use regulations such as the fair use doctrine in the United States and the GDPR in the European Union. Our experiments show that the parametric LM struggles on domains not covered by OLC. However, access to the datastore greatly improves out of domain performance, closing 90% of the performance gap with an LM trained on the Pile, a more diverse corpus with mostly high-risk text. We also analyze which nonparametric approach works best, where the remaining errors lie, and how performance scales with datastore size. Our results suggest that it is possible to build high quality language models while mitigating their legal risk.
We present a Multimodal Backdoor Defense technique TIJO (Trigger Inversion using Joint Optimization). Recent work arXiv:2112.07668 has demonstrated successful backdoor attacks on multimodal models for the Visual Question Answering task. Their dual-key backdoor trigger is split across two modalities (image and text), such that the backdoor is activated if and only if the trigger is present in both modalities. We propose TIJO that defends against dual-key attacks through a joint optimization that reverse-engineers the trigger in both the image and text modalities. This joint optimization is challenging in multimodal models due to the disconnected nature of the visual pipeline which consists of an offline feature extractor, whose output is then fused with the text using a fusion module. The key insight enabling the joint optimization in TIJO is that the trigger inversion needs to be carried out in the object detection box feature space as opposed to the pixel space. We demonstrate the effectiveness of our method on the TrojVQA benchmark, where TIJO improves upon the state-of-the-art unimodal methods from an AUC of 0.6 to 0.92 on multimodal dual-key backdoors. Furthermore, our method also improves upon the unimodal baselines on unimodal backdoors. We present ablation studies and qualitative results to provide insights into our algorithm such as the critical importance of overlaying the inverted feature triggers on all visual features during trigger inversion. The prototype implementation of TIJO is available at https://github.com/SRI-CSL/TIJO.
A consensus report produced for the Air Force Research Laboratory by the National Academies of Sciences, Engineering, and Mathematics documented a prevalent and increasing desire to support human-Artificial Intelligence (AI) teaming across military service branches. Sonalysts has begun an internal initiative to explore the training of human-AI teams. The first step in this effort is to develop a Synthetic Task Environment (STE) that is capable of facilitating research on human-AI teams. We decided to use Joint All-Domain Command and Control (JADC2) as a focus point for developing the STE because the volume of sensor inputs and decision options within the JADC2 concept likely requires the use of AI systems to enable timely decisions. Given this focus, we engaged a number of Subject-Matter Experts (SMEs) with Command and Control experience to gain insight into developing a STE that embodied the teaming challenges associated with JADC2. This report documents our initial engagement with those stakeholders. The research team identified thirteen Sonalysts employees with military backgrounds and Command and Control experience, and invited them to participate. Twelve respondents completed the survey. The team then analyzed the responses to identify themes that emerged and topics that would benefit from further analysis. The results indicated that our SMEs were amenable to research using tasks that were analogous to those encountered in military environments, as long as they required teams to process a great deal of incoming data to arrive at complex decisions. The SMEs felt that the testbed should support 'teams of teams" that represented a matrixed organization, and that it should support a robust array to spoken, text-based, and face-to-face communications.
Automatically disentangling an author's style from the content of their writing is a longstanding and possibly insurmountable problem in computational linguistics. At the same time, the availability of large text corpora furnished with author labels has recently enabled learning authorship representations in a purely data-driven manner for authorship attribution, a task that ostensibly depends to a greater extent on encoding writing style than encoding content. However, success on this surrogate task does not ensure that such representations capture writing style since authorship could also be correlated with other latent variables, such as topic. In an effort to better understand the nature of the information these representations convey, and specifically to validate the hypothesis that they chiefly encode writing style, we systematically probe these representations through a series of targeted experiments. The results of these experiments suggest that representations learned for the surrogate authorship prediction task are indeed sensitive to writing style. As a consequence, authorship representations may be expected to be robust to certain kinds of data shift, such as topic drift over time. Additionally, our findings may open the door to downstream applications that require stylistic representations, such as style transfer.
In this paper, we present "1001 Nights", an AI-native game that allows players lead in-game reality through co-created storytelling with the character driven by large language model. The concept is inspired by Wittgenstein's idea of the limits of one's world being determined by the bounds of their language. Using advanced AI tools like GPT-4 and Stable Diffusion, the second iteration of the game enables the protagonist, Shahrzad, to realize words and stories in her world. The player can steer the conversation with the AI King towards specific keywords, which then become battle equipment in the game. This blend of interactive narrative and text-to-image transformation challenges the conventional border between the game world and reality through a dual perspective. We focus on Shahrzad, who seeks to alter her fate compared to the original folklore, and the player, who collaborates with AI to craft narratives and shape the game world. We explore the technical and design elements of implementing such a game with an objective to enhance the narrative game genre with AI-generated content and to delve into AI-native gameplay possibilities.