In this paper, I introduce the retrieval problem, a simple reasoning task that can be solved only by transformers with a minimum number of layers. The task has an adjustable difficulty that can further increase the required number of layers to any arbitrary value. I demonstrate that large language models can solve the task under different prompting formulations without any fine-tuning. To understand how transformers solve the retrieval problem, I train several transformers on a minimal formulation. I find that successful learning occurs only under the presence of an implicit curriculum. I uncover the learned mechanisms by studying the attention maps in the trained transformers. I also study the training process, uncovering that attention heads always emerge in a specific sequence.

Model-Free Reinforcement Learning (RL) algorithms either learn how to map states to expected rewards or search for policies that can maximize a certain performance function. Model-Based algorithms instead, aim to learn an approximation of the underlying model of the RL environment and then use it in combination with planning algorithms. Upside-Down Reinforcement Learning (UDRL) is a novel learning paradigm that aims to learn how to predict actions from states and desired commands. This task is formulated as a Supervised Learning problem and has successfully been tackled by Neural Networks (NNs). In this paper, we investigate whether function approximation algorithms other than NNs can also be used within a UDRL framework. Our experiments, performed over several popular optimal control benchmarks, show that tree-based methods like Random Forests and Extremely Randomized Trees can perform just as well as NNs with the significant benefit of resulting in policies that are inherently more interpretable than NNs, therefore paving the way for more transparent, safe, and robust RL.

Among adversarial attacks against sequential recommender systems, model extraction attacks represent a method to attack sequential recommendation models without prior knowledge. Existing research has primarily concentrated on the adversary's execution of black-box attacks through data-free model extraction. However, a significant gap remains in the literature concerning the development of surrogate models by adversaries with access to few-shot raw data (10\% even less). That is, the challenge of how to construct a surrogate model with high functional similarity within the context of few-shot data scenarios remains an issue that requires resolution.This study addresses this gap by introducing a novel few-shot model extraction framework against sequential recommenders, which is designed to construct a superior surrogate model with the utilization of few-shot data. The proposed few-shot model extraction framework is comprised of two components: an autoregressive augmentation generation strategy and a bidirectional repair loss-facilitated model distillation procedure. Specifically, to generate synthetic data that closely approximate the distribution of raw data, autoregressive augmentation generation strategy integrates a probabilistic interaction sampler to extract inherent dependencies and a synthesis determinant signal module to characterize user behavioral patterns. Subsequently, bidirectional repair loss, which target the discrepancies between the recommendation lists, is designed as auxiliary loss to rectify erroneous predictions from surrogate models, transferring knowledge from the victim model to the surrogate model effectively. Experiments on three datasets show that the proposed few-shot model extraction framework yields superior surrogate models.

Protein language models (PLMs) are capable of learning the relationships between protein sequences and functions by treating amino acid sequences as textual data in a self-supervised manner. However, fine-tuning these models typically demands substantial computational resources and time, with results that may not always be optimized for specific tasks. To overcome these challenges, this study employs the LoRA method to perform end-to-end fine-tuning of the ESM-2 model specifically for protein property prediction tasks, utilizing only sequence information. Additionally, a multi-head attention mechanism is integrated into the downstream network to combine sequence features with contact map information, thereby enhancing the model's comprehension of protein sequences. Experimental results of extensive classification and regression tasks demonstrate that the fine-tuned model achieves strong performance and faster convergence across multiple regression and classification tasks.

Thermal comfort is essential for well-being in urban spaces, especially as cities face increasing heat from urbanization and climate change. Existing thermal comfort models usually overlook temporal dynamics alongside spatial dependencies. We address this problem by introducing a spatio-temporal jump model that clusters data with persistence across both spatial and temporal dimensions. This framework enhances interpretability, minimizes abrupt state changes, and easily handles missing data. We validate our approach through extensive simulations, demonstrating its accuracy in recovering the true underlying partition. When applied to hourly environmental data gathered from a set of weather stations located across the city of Singapore, our proposal identifies meaningful thermal comfort regimes, demonstrating its effectiveness in dynamic urban settings and suitability for real-world monitoring. The comparison of these regimes with feedback on thermal preference indicates the potential of an unsupervised approach to avoid extensive surveys.

Open-source large language models are becoming increasingly available and popular among researchers and practitioners. While significant progress has been made on open-weight models, open training data is a practice yet to be adopted by the leading open-weight models creators. At the same time, there researchers are working to make language models safer. We propose a data curation pipeline to reduce harmful outputs by models trained on public domain data. There are unique challenges to working with public domain data, as these sources differ from web text in both form and content. Many sources are historical documents and are the result of Optical Character Recognition (OCR). Consequently, current state-of-the-art approaches to toxicity filtering are often infeasible or inappropriate for open data models. In this paper, we introduce a new fully open-source pipeline for open-data toxicity filtering. Our contributions are threefold. We create a custom training dataset, ToxicCommons, which is composed of texts which have been classified across five different dimensions (racial/origin-based, gender/sex-based, religious, ability-based discrimination, and violence). We use this dataset to train a custom classifier, Celadon, that can be used to detect toxic content in open data more efficiently at a larger scale. Finally, we describe the balanced approach to content filtration that optimizes safety filtering with respect to the filtered data available for training.

Divergent brain connectivity is thought to underlie the behavioral and cognitive symptoms observed in many neurodevelopmental disorders. Quantifying divergence from neurotypical connectivity patterns offers a promising pathway to inform diagnosis and therapeutic interventions. While advanced neuroimaging techniques, such as diffusion MRI (dMRI), have facilitated the mapping of brain's structural connectome, the challenge lies in accurately modeling developmental trajectories within these complex networked structures to create robust neurodivergence markers. In this work, we present the Brain Representation via Individualized Deep Generative Embedding (BRIDGE) framework, which integrates normative modeling with a bio-inspired deep generative model to create a reference trajectory of connectivity transformation as part of neurotypical development. This will enable the assessment of neurodivergence by comparing individuals to the established neurotypical trajectory. BRIDGE provides a global neurodivergence score based on the difference between connectivity-based brain age and chronological age, along with region-wise neurodivergence maps that highlight localized connectivity differences. Application of BRIDGE to a large cohort of children with autism spectrum disorder demonstrates that the global neurodivergence score correlates with clinical assessments in autism, and the regional map offers insights into the heterogeneity at the individual level in neurodevelopmental disorders. Together, the neurodivergence score and map form powerful tools for quantifying developmental divergence in connectivity patterns, advancing the development of imaging markers for personalized diagnosis and intervention in various clinical contexts.

This paper investigates the presence of OCR-sensitive neurons within the Transformer architecture and their influence on named entity recognition (NER) performance on historical documents. By analysing neuron activation patterns in response to clean and noisy text inputs, we identify and then neutralise OCR-sensitive neurons to improve model performance. Based on two open access large language models (Llama2 and Mistral), experiments demonstrate the existence of OCR-sensitive regions and show improvements in NER performance on historical newspapers and classical commentaries, highlighting the potential of targeted neuron modulation to improve models' performance on noisy text.

Model-based reinforcement learning refers to a set of approaches capable of sample-efficient decision making, which create an explicit model of the environment. This model can subsequently be used for learning optimal policies. In this paper, we propose a temporal Gaussian Mixture Model composed of a perception model and a transition model. The perception model extracts discrete (latent) states from continuous observations using a variational Gaussian mixture likelihood. Importantly, our model constantly monitors the collected data searching for new Gaussian components, i.e., the perception model performs a form of structure learning (Smith et al., 2020; Friston et al., 2018; Neacsu et al., 2022) as it learns the number of Gaussian components in the mixture. Additionally, the transition model learns the temporal transition between consecutive time steps by taking advantage of the Dirichlet-categorical conjugacy. Both the perception and transition models are able to forget part of the data points, while integrating the information they provide within the prior, which ensure fast variational inference. Finally, decision making is performed with a variant of Q-learning which is able to learn Q-values from beliefs over states. Empirically, we have demonstrated the model's ability to learn the structure of several mazes: the model discovered the number of states and the transition probabilities between these states. Moreover, using its learned Q-values, the agent was able to successfully navigate from the starting position to the maze's exit.

Record linkage integrates diverse data sources by identifying records that refer to the same entity. In the context of mineral site records, accurate record linkage is crucial for identifying and mapping mineral deposits. Properly linking records that refer to the same mineral deposit helps define the spatial coverage of mineral areas, benefiting resource identification and site data archiving. Mineral site record linkage falls under the spatial record linkage category since the records contain information about the physical locations and non-spatial attributes in a tabular format. The task is particularly challenging due to the heterogeneity and vast scale of the data. While prior research employs pre-trained discriminative language models (PLMs) on spatial entity linkage, they often require substantial amounts of curated ground-truth data for fine-tuning. Gathering and creating ground truth data is both time-consuming and costly. Therefore, such approaches are not always feasible in real-world scenarios where gold-standard data are unavailable. Although large generative language models (LLMs) have shown promising results in various natural language processing tasks, including record linkage, their high inference time and resource demand present challenges. We propose a method that leverages an LLM to generate training data and fine-tune a PLM to address the training data gap while preserving the efficiency of PLMs. Our approach achieves over 45\% improvement in F1 score for record linkage compared to traditional PLM-based methods using ground truth data while reducing the inference time by nearly 18 times compared to relying on LLMs. Additionally, we offer an automated pipeline that eliminates the need for human intervention, highlighting this approach's potential to overcome record linkage challenges.