We propose Universal Document Processing (UDOP), a foundation Document AI model which unifies text, image, and layout modalities together with varied task formats, including document understanding and generation. UDOP leverages the spatial correlation between textual content and document image to model image, text, and layout modalities with one uniform representation. With a novel Vision-Text-Layout Transformer, UDOP unifies pretraining and multi-domain downstream tasks into a prompt-based sequence generation scheme. UDOP is pretrained on both large-scale unlabeled document corpora using innovative self-supervised objectives and diverse labeled data. UDOP also learns to generate document images from text and layout modalities via masked image reconstruction. To the best of our knowledge, this is the first time in the field of document AI that one model simultaneously achieves high-quality neural document editing and content customization. Our method sets the state-of-the-art on 8 Document AI tasks, e.g., document understanding and QA, across diverse data domains like finance reports, academic papers, and websites. UDOP ranks first on the leaderboard of the Document Understanding Benchmark.

Our concern is the problem of determining the data complexity of answering an ontology-mediated query (OMQ) formulated in linear temporal logic LTL over (Z,<) and deciding whether it is rewritable to an FO(<)-query, possibly with some extra predicates. First, we observe that, in line with the circuit complexity and FO-definability of regular languages, OMQ answering in AC0, ACC0 and NC1 coincides with FO(<,≡)-rewritability using unary predicates x ≡ 0 (mod n), FO(<,MOD)-rewritability, and FO(RPR)-rewritability using relational primitive recursion, respectively. We prove that, similarly to known PSᴘᴀᴄᴇ-completeness of recognising FO(<)-definability of regular languages, deciding FO(<,≡)- and FO(<,MOD)-definability is also PSᴘᴀᴄᴇ-complete (unless ACC0 = NC1). We then use this result to show that deciding FO(<)-, FO(<,≡)- and FO(<,MOD)-rewritability of LTL OMQs is ExᴘSᴘᴀᴄᴇ-complete, and that these problems become PSᴘᴀᴄᴇ-complete for OMQs with a linear Horn ontology and an atomic query, and also a positive query in the cases of FO(<)- and FO(<,≡)-rewritability. Further, we consider FO(<)-rewritability of OMQs with a binary-clause ontology and identify OMQ classes, for which deciding it is PSᴘᴀᴄᴇ-, Π2p- and coNP-complete.

The Targeted Maximum Likelihood Estimation (TMLE) statistical data analysis framework integrates machine learning, statistical theory, and statistical inference to provide a least biased, efficient and robust strategy for estimation and inference of a variety of statistical and causal parameters. We describe and evaluate the epidemiological applications that have benefited from recent methodological developments. We conducted a systematic literature review in PubMed for articles that applied any form of TMLE in observational studies. We summarised the epidemiological discipline, geographical location, expertise of the authors, and TMLE methods over time. We used the Roadmap of Targeted Learning and Causal Inference to extract key methodological aspects of the publications. We showcase the contributions to the literature of these TMLE results. Of the 81 publications included, 25% originated from the University of California at Berkeley, where the framework was first developed by Professor Mark van der Laan. By the first half of 2022, 70% of the publications originated from outside the United States and explored up to 7 different epidemiological disciplines in 2021-22. Double-robustness, bias reduction and model misspecification were the main motivations that drew researchers towards the TMLE framework. Through time, a wide variety of methodological, tutorial and software-specific articles were cited, owing to the constant growth of methodological developments around TMLE. There is a clear dissemination trend of the TMLE framework to various epidemiological disciplines and to increasing numbers of geographical areas. The availability of R packages, publication of tutorial papers, and involvement of methodological experts in applied publications have contributed to an exponential increase in the number of studies that understood the benefits, and adoption, of TMLE.

Identifying words that impact a task's performance more than others is a challenge in natural language processing. Transformers models have recently addressed this issue by incorporating an attention mechanism that assigns greater attention (i.e., relevance) scores to some words than others. Because of the attention mechanism's high computational cost, transformer models usually have an input-length limitation caused by hardware constraints. This limitation applies to many transformers, including the well-known bidirectional encoder representations of the transformer (BERT) model. In this paper, we examined BERT's attention assignment mechanism, focusing on two questions: (1) How can attention be employed to reduce input length? (2) How can attention be used as a control mechanism for conditional text generation? We investigated these questions in the context of a text classification task. We discovered that BERT's early layers assign more critical attention scores for text classification tasks compared to later layers. We demonstrated that the first layer's attention sums could be used to filter tokens in a given sequence, considerably decreasing the input length while maintaining good test accuracy. We also applied filtering, which uses a compute-efficient semantic similarities algorithm, and discovered that retaining approximately 6\% of the original sequence is sufficient to obtain 86.5\% accuracy. Finally, we showed that we could generate data in a stable manner and indistinguishable from the original one by only using a small percentage (10\%) of the tokens with high attention scores according to BERT's first layer.

Nowadays, numerous applications incorporate machine learning (ML) algorithms due to their prominent achievements. However, many studies in the field of computer vision have shown that ML can be fooled by intentionally crafted instances, called adversarial examples. These adversarial examples take advantage of the intrinsic vulnerability of ML models. Recent research raises many concerns in the cybersecurity field. An increasing number of researchers are studying the feasibility of such attacks on security systems based on ML algorithms, such as Intrusion Detection Systems (IDS). The feasibility of such adversarial attacks would be influenced by various domain-specific constraints. This can potentially increase the difficulty of crafting adversarial examples. Despite the considerable amount of research that has been done in this area, much of it focuses on showing that it is possible to fool a model using features extracted from the raw data but does not address the practical side, i.e., the reverse transformation from theory to practice. For this reason, we propose a review browsing through various important papers to provide a comprehensive analysis. Our analysis highlights some challenges that have not been addressed in the reviewed papers.

Niki Parmar and Ashish Vaswani co-authored a seminal paper that set the groundwork for ChatGPT and other generative AI models. ChatGPT has taken the world by storm, but seeds of the groundbreaking technology were sown at the USC Viterbi School of Engineering. The seminal paper "Attention Is All You Need," which laid the foundation for ChatGPT and other generative AI systems, was co-authored by Ashish Vaswani, a PhD computer science graduate ('14) and Niki Parmar, a master's in computer science graduate ('15). The landmark paper was presented at the 2017 Conference on Neural Information Processing Systems (NeurIPS), one of the top conferences in AI and machine learning. In the paper, the researchers introduced the transformer architecture, a powerful type of neural network that has become widely used for natural language processing tasks, from text classification to language modeling.

In recent years, learning-based approaches have revolutionized motion planning. The data generation process for these methods involves caching a large number of high quality paths for different queries (start, goal pairs) in various environments. Conventionally, a uniform random strategy is used for sampling these queries. However, this leads to inclusion of "trivial paths" in the dataset (e.g.,, straight line paths in case of length-optimal planning), which can be solved efficiently if the planner has access to a steering function. This work proposes a "non-trivial" query sampling procedure to add more complex paths in the dataset. Numerical experiments show that a higher success rate can be attained for neural planners trained on such a non-trivial dataset.

In conceptual modeling (CM), humans apply abstraction to represent excerpts of reality for means of understanding and communication, and processing by machines. Artificial Intelligence (AI) is applied to vast amounts of data to automatically identify patterns or classify entities. While CM produces comprehensible and explicit knowledge representations, the outcome of AI algorithms often lacks these qualities while being able to extract knowledge from large and unstructured representations. Recently, a trend toward intertwining CM and AI emerged. This systematic mapping study shows how this interdisciplinary research field is structured, which mutual benefits are gained by the intertwining, and future research directions.

With the increasing complexity of robot systems, it is necessary to simulate them before deployment. To do this, a model of the robot's kinematics or dynamics is required. One of the most commonly used formats for modeling robots is the Unified Robot Description Format (URDF). The goal of this article is to understand how URDF is currently used, what challenges people face when working with it, and how the community sees the future of URDF. The outcome can potentially be used to guide future research. This article presents the results from a survey based on 510 anonymous responses from robotic developers of different backgrounds and levels of experience. We find that 96.8% of the participants have simulated robots before, and of them 95.5% had used URDF. We identify a number of challenges and limitations that complicate the use of URDF, such as the inability to model parallel linkages and closed-chain systems, no real standard, lack of documentation, and a limited number of dynamic parameters to model the robot. Future perspectives for URDF are also determined, where 53.5% believe URDF will be more commonly used in the future, 12.2% believe other standards or tools will make URDF obsolete, and 34.4% are not sure what the future of URDF will be. Most participants agree that there is a need for better tooling to ensure URDF's future use.

We present a novel dataset collected by ASOS (a major online fashion retailer) to address the challenge of predicting customer returns in a fashion retail ecosystem. With the release of this substantial dataset we hope to motivate further collaboration between research communities and the fashion industry. We first explore the structure of this dataset with a focus on the application of Graph Representation Learning in order to exploit the natural data structure and provide statistical insights into particular features within the data. In addition to this, we show examples of a return prediction classification task with a selection of baseline models (i.e. with no intermediate representation learning step) and a graph representation based model. We show that in a downstream return prediction classification task, an F1-score of 0.792 can be found using a Graph Neural Network (GNN), improving upon other models discussed in this work. Alongside this increased F1-score, we also present a lower cross-entropy loss by recasting the data into a graph structure, indicating more robust predictions from a GNN based solution. These results provide evidence that GNNs could provide more impactful and usable classifications than other baseline models on the presented dataset and with this motivation, we hope to encourage further research into graph-based approaches using the ASOS GraphReturns dataset.