Publications
Our teams aspire to make discoveries that impact everyone, and core to our approach is sharing our research and tools to fuel progress in the field.
Our teams aspire to make discoveries that impact everyone, and core to our approach is sharing our research and tools to fuel progress in the field.
Sort By
1 - 15 of 10129 publications
Demystifying Embedding Spaces using Large Language Models
Jihwan Jeong
Lior Shani
Martin Mladenov
The Twelfth International Conference on Learning Representations (2024)
Preview abstract
Embeddings have become a pivotal means to represent complex, multi-faceted information about entities, concepts, and relationships in a condensed and useful format. Nevertheless, they often preclude direct interpretation. While downstream tasks make use of these compressed representations, meaningful interpretation usually requires visualization using dimensionality reduction or specialized machine learning interpretability methods. This paper addresses the challenge of making such embeddings more interpretable and broadly useful, by employing large language models (LLMs) to directly interact with embeddings -- transforming abstract vectors into understandable narratives. By injecting embeddings into LLMs, we enable querying and exploration of complex embedding data. We demonstrate our approach on a variety of diverse tasks, including: enhancing concept activation vectors (CAVs), communicating novel embedded entities, and decoding user preferences in recommender systems. Our work couples the immense information potential of embeddings with the interpretative power of LLMs.
View details
PaLI-X: On Scaling up a Multilingual Vision and Language Model
Josip Djolonga
Piotr Padlewski
Basil Mustafa
Carlos Riquelme
Sebastian Goodman
Yi Tay
Siamak Shakeri
Daniel Salz
Michael Tschannen
Mandar Joshi
Filip Pavetić
Gang Li
Anurag Arnab
Yuanzhong Xu
Keran Rong
Neil Houlsby
Computer Vision and Pattern Recognition Conference (CVPR) (2024)
Preview abstract
We explore the boundaries of scaling up a multilingual vision and language model, both in terms of size of the components and the breadth of its training task mixture. Our model achieves new levels of performance on a wide-range of varied and complex tasks, including multiple image-based captioning and question-answering tasks, image-based document understanding and few-shot (in-context) learning, as well as object detection, video question answering, and video captioning. Our model advances the state-of-the-art on most vision-and-language benchmarks considered (20+ of them). Finally, we observe emerging capabilities, such as complex counting and multilingual object detection, tasks that are not explicitly in the training mix.
View details
LMDX: Language Model-based Document Information Extraction And Localization
Kai Kang
Florian Luisier
Xiaoyu Sun
Ramya Sree Boppana
Zilong Wang
Jiaqi Mu
Hao Zhang
Nan Hua
Findings of the Association for Computational Linguistics ACL 2024, Association for Computational Linguistics, Bangkok, Thailand and virtual meeting, pp. 15140-15168
Preview abstract
Large Language Models (LLM) have revolutionized Natural Language Processing (NLP), improving state-of-the-art and exhibiting emergent capabilities across various tasks. However, their application in extracting information from visually rich documents, which is at the core of many document processing workflows and involving the extraction of key entities from semi-structured documents, has not yet been successful. The main obstacles to adopting LLMs for this task include the absence of layout encoding within LLMs, which is critical for high quality extraction, and the lack of a grounding mechanism to localize the predicted entities within the document. In this paper, we introduce Language Model-based Document Information EXtraction and Localization (LMDX), a methodology to reframe the document information extraction task for a LLM. LMDX enables extraction of singular, repeated, and hierarchical entities, both with and without training data, while providing grounding guarantees and localizing the entities within the document. Finally, we apply LMDX to the PaLM 2-S and Gemini Pro LLMs and evaluate it on VRDU and CORD benchmarks, setting a new state-of-the-art and showing how LMDX enables the creation of high quality, data-efficient parsers.
View details
Preview abstract
Inter-sentence pauses are the silences that occur between sentences in a paragraph or a dialogue.
They are an important aspect of long-form speech prosody, as they can affect the naturalness, intelligibility, and effectiveness of communication.
However, the user perception of inter-sentence pauses in long-form speech synthesis is not well understood. Previous work often evaluates pause modelling in conjunction with other prosodic features making it hard to explicitly study how raters perceive differences in inter-sentence pause lengths.
In this paper, using multiple text-to-speech (TTS) datasets that cover different content types, domains, and settings, we investigate how sensitive raters are to changes to the durations of inter-sentence pauses in long-form speech by comparing ground truth audio samples with renditions that have manipulated pause durations.
This experimental design is meant to allow us to draw conclusions regarding the utility that can be expected from similar evaluations when applied to synthesized long-form speech.
We find that, using standard evaluation methodologies, raters are not sensitive to variations in pause lengths unless these deviate exceedingly from the norms or expectations of the speech context.
View details
Preview abstract
As AI systems quickly improve in both breadth and depth of performance, they lend themselves to creating increasingly powerful and realistic agents, including the possibility of agents modeled on specific people. We anticipate that within our lifetimes it may become common practice for people to create a custom AI agent to interact with loved ones and/or the broader world after death. We call these generative ghosts, since such agents will be capable of generating novel content rather than merely parroting content produced by their creator while living. In this paper, we first discuss the design space of potential implementations of generative ghosts. We then discuss the practical and ethical implications of generative ghosts, including potential positive and negative impacts on individuals and society. Based on these considerations, we lay out a research agenda for the AI and HCI research communities to empower people to create and interact with AI afterlives in a safe and beneficial manner.
View details
Preview abstract
In this paper, we introduce DiarizationLM, a framework to leverage large language models (LLM) to post-process the outputs from a speaker diarization system. Various goals can be achieved with the proposed framework, such as improving the readability of the diarized transcript, or reducing the word diarization error rate (WDER). In this framework, the outputs of the automatic speech recognition (ASR) and speaker diarization systems are represented as a compact textual format, which is included in the prompt to an optionally finetuned LLM. The outputs of the LLM can be used as the refined diarization results with the desired enhancement. As a post-processing step, this framework can be easily applied to any off-the-shelf ASR and speaker diarization systems without retraining existing components. Our experiments show that a finetuned PaLM 2-S model can reduce the WDER by rel. 55.5% on the Fisher telephone conversation dataset, and rel. 44.9% on the Callhome English dataset.
View details
Preview abstract
Structured Complex Task Decomposition (SCTD) is the problem of breaking down a complex real-world task (such as planning a wedding) into a directed acyclic graph over individual steps that contribute to achieving the task, with edges specifying temporal dependencies between them. SCTD is an important component of assistive planning tools, and a challenge for commonsense reasoning systems. We probe how accurately SCTD can be done with the knowledge extracted from Large Language Models (LLMs). We introduce a high-quality human-annotated dataset for this problem and novel metrics to fairly assess performance of LLMs against several baselines. Our experiments reveal that LLMs are able to decompose complex tasks into individual steps effectively, with a relative improvement of 15% to 280% over the best baseline. We also propose a number of approaches to further improve their performance, with a relative improvement of 7% to 37% over the base model. However, we find that LLMs still struggle to predict pairwise temporal dependencies, which reveals a gap in their understanding of complex tasks.
View details
ASTRA-5G: Automated Over-the-Air Security Testing and Research Architecture for 5G SA Devices
Aanjhan Ranganathan
Christina Pöpper
Evangelos Bitsikas
Michele Guerra
Roger Piqueras Jover
Syed Khandker
WiSec '24: Proceedings of the 17th ACM Conference on Security and Privacy in Wireless and Mobile Networks, ACM (2024)
Preview abstract
Despite the widespread deployment of 5G technologies, there exists a critical gap in security testing for 5G Standalone (SA) devices. Existing methods, largely manual and labor-intensive, are ill-equipped to fully uncover the state of security in the implementations of 5G-SA protocols and standards on devices, severely limiting the ability to conduct comprehensive evaluations. To address this issue, in this work, we introduce an novel, open-source framework that auto-
mates the security testing process for 5G SA devices. By leveraging enhanced functionalities of 5G SA core and Radio Access Network (RAN) software, our framework offers a streamlined approach to generating, executing, and evaluating test cases, specifically focusing on the Non-Access Stratum (NAS) layer. Our application of this framework across multiple 5G SA devices provides in-depth security insights, significantly improving testing efficiency and breadth.
View details
Minimizing Live Experiments in Recommender Systems: User Simulation to Evaluate Preference Elicitation Policies
Martin Mladenov
James Pine
Hubert Pham
Shane Li
Xujian Liang
Anton Polishko
Li Yang
Ben Scheetz
Proceedings of he 47th International ACM/SIGIR Conference on Research and Development in Information Retrieval (SIGIR-24), Washington, DC (2024), pp. 2925-2929
Preview abstract
Evaluation of policies in recommender systems (RSs) typically involves A/B testing using live experiments on real users to assess a new policy's impact on relevant metrics. This ``gold standard'' comes at a high cost, however, in terms of cycle time, user cost, and potential user retention. In developing policies for onboarding new users, these costs can be especially problematic, since on-boarding occurs only once. In this work, we describe a simulation methodology used to augment (and reduce) the use of live experiments. We illustrate its deployment for the evaluation of preference elicitation algorithms used to onboard new users of the YouTube Music platform. By developing counterfactually robust user behavior models, and a simulation service that couples such models with production infrastructure, we are able to test new algorithms in a way that reliably predicts their performance on key metrics when deployed live, sometimes more reliably than live experiments due to the scale at which simulation can be realized. We describe our domain, our simulation models and platform, results of experiments and deployment, and suggest future steps needed to further realistic simulation as a powerful complement to live experiments.
View details
A versatile, semi-automated image analysis workflow for time-lapse camera trap image classification
Hanna Böhner
Olga Pokrovskaya
Desheng Liu
Natalia Sokolova
Olivier Gilg
Wenbo Zhou
Ivan Fufachev
Peter Ungar
Rolf Anker Ims
Alexsandr Sokolov
Dorothee Ehrich
Gerardo Celis
Ecological Informatics (2024)
Preview abstract
Camera traps are a powerful, practical, and non-invasive method used widely to monitor animal communities and evaluate management actions. However, camera trap arrays can generate thousands to millions of images that require significant time and effort to review. Computer vision has emerged as a tool to accelerate this image review process. We propose a multi-step, semi-automated workflow which takes advantage of site-specific and generalizable models to improve detections and consists of (1) automatically identifying and removing low-quality images in parallel with classification into animals, humans, vehicles, and empty, (2) automatically cropping objects from images and classifying them (rock, bait, empty, and species), and (3) manually inspecting a subset of images. We trained and evaluated this approach using 548,627 images from 46 cameras in two regions of the Arctic: “Finnmark” (Finnmark County, Norway) and “Yamal” (Yamalo-Nenets Autonomous District, Russia). The automated steps yield image classification accuracies of 92% and 90% for the Finnmark and Yamal sets, respectively, reducing the number of images that required manual inspection to 9.2% of the Finnmark set and 3.9% of the Yamal set. The amount of time invested in developing models would be offset by the time saved from automation in about three seasons/years. Researchers can modify this multi-step process to develop their own site-specific models and meet other needs for monitoring and surveying wildlife, balancing the acceptable levels of false negatives and positives.
View details
Validation of a deep learning system for the detection of diabetic retinopathy in Indigenous Australians
Mark Chia
Fred Hersch
Pearse Keane
Angus Turner
British Journal of Ophthalmology, 108 (2024), pp. 268-273
Preview abstract
Background/aims: Deep learning systems (DLSs) for diabetic retinopathy (DR) detection show promising results but can underperform in racial and ethnic minority groups, therefore external validation within these populations is critical for health equity. This study evaluates the performance of a DLS for DR detection among Indigenous Australians, an understudied ethnic group who suffer disproportionately from DR-related blindness.
Methods: We performed a retrospective external validation study comparing the performance of a DLS against a retinal specialist for the detection of more-than-mild DR (mtmDR), vision-threatening DR (vtDR) and all-cause referable DR. The validation set consisted of 1682 consecutive, single-field, macula-centred retinal photographs from 864 patients with diabetes (mean age 54.9 years, 52.4% women) at an Indigenous primary care service in Perth, Australia. Three-person adjudication by a panel of specialists served as the reference standard.
Results: For mtmDR detection, sensitivity of the DLS was superior to the retina specialist (98.0% (95% CI, 96.5 to 99.4) vs 87.1% (95% CI, 83.6 to 90.6), McNemar’s test p<0.001) with a small reduction in specificity (95.1% (95% CI, 93.6 to 96.4) vs 97.0% (95% CI, 95.9 to 98.0), p=0.006). For vtDR, the DLS’s sensitivity was again superior to the human grader (96.2% (95% CI, 93.4 to 98.6) vs 84.4% (95% CI, 79.7 to 89.2), p<0.001) with a slight drop in specificity (95.8% (95% CI, 94.6 to 96.9) vs 97.8% (95% CI, 96.9 to 98.6), p=0.002). For all-cause referable DR, there was a substantial increase in sensitivity (93.7% (95% CI, 91.8 to 95.5) vs 74.4% (95% CI, 71.1 to 77.5), p<0.001) and a smaller reduction in specificity (91.7% (95% CI, 90.0 to 93.3) vs 96.3% (95% CI, 95.2 to 97.4), p<0.001).
Conclusion: The DLS showed improved sensitivity and similar specificity compared with a retina specialist for DR detection. This demonstrates its potential to support DR screening among Indigenous Australians, an underserved population with a high burden of diabetic eye disease.
View details
Preview abstract
Interruptions in digital services are a common occurrence for users. These disruptions, however, exact a cost in terms of attention, task completion rate, and, most importantly, emotional state. While several methods currently employed by service providers attempt to address this, the paper will argue that browser games or similar interactive interfaces should become a standard mechanism to ease the aforementioned effects.
View details
Preview abstract
Interactions with Extended Reality Head Mounted Devices (XR HMDs) applications require precise, intuitive and efficient input methods. Current approaches either rely on power-intensive sensors, such as cameras for hand-tracking, or specialized hardware in the form of handheld controllers. As an alternative, past works have explored the use of devices already present with the user, in the form of smartphones and smartwatches as practical input solutions. However, this approach risks interaction overload---how can one determine whether the user’s interaction gestures on the watch-face or phone screen are directed toward control of the mobile device itself or the XR device? To this effect, we propose a novel framework for cross-device input routing and device arbitration by employing Inertial Measurement Units (IMUs) within these devices. We validate our approach in a user study with six participants. By making use of the relative orientation between the headset and the target input device, we can estimate the intended device of interaction with 93.7% accuracy. Our method offers a seamless, energy-efficient alternative for input management in XR, enhancing user experience through natural and ergonomic interactions.
View details
Websites Need Your Permission Too – User Sentiment and Decision Making on Web Permission Prompts in Desktop Chrome
Marian Harbach
CHI 2024, ACM
Preview abstract
The web utilizes permission prompts to moderate access to certain capabilities. We present the first investigation of user behavior and sentiment of this security and privacy measure on the web, using 28 days of telemetry data from more than 100M Chrome installations on desktop platforms and experience sampling responses from 25,706 Chrome users. Based on this data, we find that ignoring and dismissing permission prompts are most common for geolocation and notifications. Permission prompts are perceived as more annoying and interrupting when they are not allowed, and most respondents cite a rational reason for the decision they took. Our data also supports that the perceived availability of contextual information from the requesting website is associated with allowing access to a requested capability. More usable permission controls could facilitate adoption of best practices that address several of the identified challenges; and ultimately could lead to better user experiences and a safer web.
View details
Promises and Pitfalls of Generative Masked Language Modeling: Theoretical Framework and Practical Guidelines
Yuchen Li
Alexandre Kirchmeyer
Aashay Mehta
Yilong Qin
Andrej Risteski
International Conference on Machine Learning (2024) (to appear)
Preview abstract
Autoregressive language models are the currently dominant paradigm for text generation, however they have some fundamental limitations that cannot be remedied by scale ---for example inherently sequential and unidirectional generation. While alternate classes of models have been explored, we have limited mathematical understanding of their fundamental power and limitations. In this paper we focus on Generative Masked Language Models (GMLMs), a non-autoregressive paradigm in which we train a model to fit conditional probabilities of the data distribution via masking, which are subsequently used as inputs to a Markov Chain to draw samples from the model. These models empirically strike a promising speed-quality trade-off as each step can be typically parallelized by decoding the entire sequence in parallel. We develop a mathematical framework for analyzing and improving such models which sheds light on questions of sample complexity and inference speed and quality. Empirically, we adapt the T5 model for iteratively-refined parallel decoding, achieving 2-3x speedup in machine translation with minimal sacrifice in quality compared with autoregressive models. We run careful ablation experiments to give recommendations on key design choices, and make fine-grained observations on the common error modes in connection with our theory. Our mathematical analyses and empirical observations characterize both potentials and limitations of this approach, and can be applied to future works on improving understanding and performance of GMLMs.
View details
