Text normalization is a crucial technology for low-resource languages which lack rigid spelling conventions or that have undergone multiple spelling reforms. Low-resource text normalization has so far relied upon hand-crafted rules, which are perceived to be more data efficient than neural methods. In this paper we examine the case of text normalization for Ligurian, an endangered Romance language. We collect 4,394 Ligurian sentences paired with their normalized versions, as well as the first open source monolingual corpus for Ligurian. We show that, in spite of the small amounts of data available, a compact transformer-based model can be trained to achieve very low error rates by the use of backtranslation and appropriate tokenization.

Recent generative AI platforms are able to create texts or impressive images from simple text prompts. This makes them powerful tools for summarizing knowledge about architectural history or deriving new creative work in early design tasks like ideation, sketching and modelling. But, how good is the understanding of the generative AI models of the history of architecture? Has it learned to properly distinguish styles, or is it hallucinating information? In this chapter, we investigate this question for generative AI platforms for text and image generation for different architectural styles, to understand the capabilities and boundaries of knowledge of those tools. We also analyze how they are already being used by analyzing a data set of 101 million Midjourney queries to see if and how practitioners are already querying for specific architectural concepts.

Neural networks excel at discovering statistical patterns in high-dimensional data sets. In practice, higher-order cumulants, which quantify the non-Gaussian correlations between three or more variables, are particularly important for the performance of neural networks. But how efficient are neural networks at extracting features from higher-order cumulants? We study this question in the spiked cumulant model, where the statistician needs to recover a privileged direction or "spike" from the order-$p\ge 4$ cumulants of~$d$-dimensional inputs. We first characterise the fundamental statistical and computational limits of recovering the spike by analysing the number of samples~$n$ required to strongly distinguish between inputs from the spiked cumulant model and isotropic Gaussian inputs. We find that statistical distinguishability requires $n\gtrsim d$ samples, while distinguishing the two distributions in polynomial time requires $n \gtrsim d^2$ samples for a wide class of algorithms, i.e. those covered by the low-degree conjecture. These results suggest the existence of a wide statistical-to-computational gap in this problem. Numerical experiments show that neural networks learn to distinguish the two distributions with quadratic sample complexity, while "lazy" methods like random features are not better than random guessing in this regime. Our results show that neural networks extract information from higher-order correlations in the spiked cumulant model efficiently, and reveal a large gap in the amount of data required by neural networks and random features to learn from higher-order cumulants.

In recent years, there has been a significant transformation in medical research methodologies towards the adoption of Deep Learning (DL) techniques for predicting critical events, such as disease development and patient mortality. Despite their potential to handle complex data, practical applications in this domain remain limited, with most studies still relying on traditional statistical methods. Survival Analysis (SA), or time-to-event analysis, is an essential tool for studying specific events in various disciplines, not only in medicine but also in fields such as recommendation systems [1], employee retention [2], market modeling [3], and financial risk assessment [4]. According to the existing literature, the Cox proportional hazards model (Cox-PH) [5] is the dominant SA method that offers a semiparametric regression solution to the non-parametric Kaplan-Meier estimator problem [6]. Unlike the Kaplan-Meier method, which uses a single covariate, Cox-PH incorporates multiple covariates to predict event times and assess their impact on the hazard rate at specific time points. However, it is crucial to acknowledge that the Cox-PH model is built on certain strong assumptions. One of these is the proportional hazards assumption, which posits that different individuals have hazard functions that remain constant over time. Furthermore, the model assumes a linear relation between the natural logarithm of the relative hazard (the ratio of the hazard at time t to the baseline hazard) and the covariates. Furthermore, it assumes the absence of interactions among these covariates.

Understanding user intentions is crucial for enhancing product recommendations, navigation suggestions, and query reformulations. However, user intentions can be complex, involving multiple sessions and attribute requirements connected by logical operators such as And, Or, and Not. For example, a user may search for Nike or Adidas running shoes across various sessions, with a preference for the color purple. In another case, a user may have purchased a mattress in a previous session and is now seeking a corresponding bed frame without intending to buy another mattress. Prior research on session understanding has not sufficiently addressed how to make product or attribute recommendations for such complex intentions. In this paper, we introduce the task of logical session complex query answering, where sessions are treated as hyperedges of items, and we formulate the problem of complex intention understanding as a task of logical session complex queries answering (LS-CQA) on an aggregated hypergraph of sessions, items, and attributes. The proposed task is a special type of complex query answering task with sessions as ordered hyperedges. We also propose a new model, the Logical Session Graph Transformer (LSGT), which captures interactions among items across different sessions and their logical connections using a transformer structure. We analyze the expressiveness of LSGT and prove the permutation invariance of the inputs for the logical operators. We evaluate LSGT on three datasets and demonstrate that it achieves state-of-the-art results.

Large language models (LLMs) are able to engage in natural-sounding conversations with humans, showcasing unprecedented capabilities for information retrieval and automated decision support. They have disrupted human-technology interaction and the way businesses operate. However, technologies based on generative artificial intelligence (GenAI) are known to hallucinate, misinform, and display biases introduced by the massive datasets on which they are trained. Existing research indicates that humans may unconsciously internalize these biases, which can persist even after they stop using the programs. This study explores the cultural self-perception of LLMs by prompting ChatGPT (OpenAI) and Bard (Google) with value questions derived from the GLOBE project. The findings reveal that their cultural self-perception is most closely aligned with the values of English-speaking countries and countries characterized by sustained economic competitiveness. Recognizing the cultural biases of LLMs and understanding how they work is crucial for all members of society because one does not want the black box of artificial intelligence to perpetuate bias in humans, who might, in turn, inadvertently create and train even more biased algorithms.

Tensorial neural networks (TNNs) combine the successes of multilinear algebra with those of deep learning to enable extremely efficient reduced-order models of high-dimensional problems. Here, I describe a deep neural network architecture that fuses multiple TNNs into a larger network, intended to solve a broader class of problems than a single TNN. I evaluate this architecture, referred to as a "stacked tensorial neural network" (STNN), on a parametric PDE with three independent variables and three parameters. The three parameters correspond to one PDE coefficient and two quantities describing the domain geometry. The STNN provides an accurate reduced-order description of the solution manifold over a wide range of parameters. There is also evidence of meaningful generalization to parameter values outside its training data. Finally, while the STNN architecture is relatively simple and problem agnostic, it can be regularized to incorporate problem-specific features like symmetries and physical modeling assumptions.

Social recommendations have been widely adopted in substantial domains. Recently, graph neural networks (GNN) have been employed in recommender systems due to their success in graph representation learning. However, dealing with the dynamic property of social network data is a challenge. This research presents a novel method that provides social recommendations by incorporating the dynamic property of social network data in a heterogeneous graph. The model aims to capture user preference over time without going through the complexities of a dynamic graph by adding period nodes to define users' long-term and short-term preferences and aggregating assigned edge weights. The model is applied to real-world data to argue its superior performance. Promising results demonstrate the effectiveness of this model.

Language model attacks typically assume one of two extreme threat models: full white-box access to model weights, or black-box access limited to a text generation API. However, real-world APIs are often more flexible than just text generation: these APIs expose ``gray-box'' access leading to new threat vectors. To explore this, we red-team three new functionalities exposed in the GPT-4 APIs: fine-tuning, function calling and knowledge retrieval. We find that fine-tuning a model on as few as 15 harmful examples or 100 benign examples can remove core safeguards from GPT-4, enabling a range of harmful outputs. Furthermore, we find that GPT-4 Assistants readily divulge the function call schema and can be made to execute arbitrary function calls. Finally, we find that knowledge retrieval can be hijacked by injecting instructions into retrieval documents. These vulnerabilities highlight that any additions to the functionality exposed by an API can create new vulnerabilities.

Despite the impressive advancements achieved using deep-learning for functional brain activity analysis, the heterogeneity of functional patterns and scarcity of imaging data still pose challenges in tasks such as prediction of future onset of Post-Traumatic Epilepsy (PTE) from data acquired shortly after traumatic brain injury (TBI). Foundation models pre-trained on separate large-scale datasets can improve the performance from scarce and heterogeneous datasets. For functional Magnetic Resonance Imaging (fMRI), while data may be abundantly available from healthy controls, clinical data is often scarce, limiting the ability of foundation models to identify clinically-relevant features. We overcome this limitation by introducing a novel training strategy for our foundation model by integrating meta-learning with self-supervised learning to improve the generalization from normal to clinical features. In this way we enable generalization to other downstream clinical tasks, in our case prediction of PTE. To achieve this, we perform self-supervised training on the control dataset to focus on inherent features that are not limited to a particular supervised task while applying meta-learning, which strongly improves the model's generalizability using bi-level optimization. Through experiments on neurological disorder classification tasks, we demonstrate that the proposed strategy significantly improves task performance on small-scale clinical datasets. To explore the generalizability of the foundation model in downstream applications, we then apply the model to an unseen TBI dataset for prediction of PTE using zero-shot learning. Results further demonstrated the enhanced generalizability of our foundation model.