Text classification, a classic task in natural language processing (NLP), involves assigning predefined categories to textual data and is crucial for applications ranging from sentiment analysis to spam detection. This thesis advances text classification by harnessing the intrinsic knowledge of Pretrained Language Models (PLMs) to address three challenging scenarios: distractor selection for multiple-choice cloze questions, improving robustness for prompt-based zero-shot text classification, and demonstration selection for retrieval-based in-context learning. Firstly, we focus on selecting distractors for multiple-choice cloze questions, ensuring that they are misleading yet incorrect. We assess the relationship between human experts' annotations (accept/reject) and various features, including context-free features (e.g., word frequency) and context-sensitive features (e.g., conditional probabilities of fillin-the-blank words). We utilize pretrained embeddings and follow annotation instructions for context-free feature design, and we find that using contextualized word representations from PLMs as features drastically improves performance over traditional feature-based models, even rivaling human performance (Chapter 3).

The widespread adoption of cloud-based proprietary large language models (LLMs) has introduced significant challenges, including operational dependencies, privacy concerns, and the necessity of continuous internet connectivity. In this work, we introduce an LLMOps pipeline, "LlamaDuo", for the seamless migration of knowledge and abilities from service-oriented LLMs to smaller, locally manageable models. This pipeline is crucial for ensuring service continuity in the presence of operational failures, strict privacy policies, or offline requirements. Our LlamaDuo involves fine-tuning a small language model against the service LLM using a synthetic dataset generated by the latter. If the performance of the fine-tuned model falls short of expectations, it is enhanced by further fine-tuning with additional similar data created by the service LLM. This iterative process guarantees that the smaller model can eventually match or even surpass the service LLM's capabilities in specific downstream tasks, offering a practical and scalable solution for managing AI deployments in constrained environments. Extensive experiments with leading-edge LLMs are conducted to demonstrate the effectiveness, adaptability, and affordability of LlamaDuo across various downstream tasks.

Temporal Point Processes (TPPs) have recently become increasingly interesting for learning dynamics in graph data. A reason for this is that learning on dynamic graph data is becoming more relevant, since data from many scientific fields, ranging from mathematics, biology, social sciences, and physics to computer science, is naturally related and inherently dynamic. In addition, TPPs provide a meaningful characterization of event streams and a prediction mechanism for future events. Therefore, (semi-)parameterized Neural TPPs have been introduced whose characterization can be (partially) learned and, thus, enable the representation of more complex phenomena. However, the research on modeling dynamic graphs with TPPs is relatively young, and only a few models for node attribute changes or evolving edges have been proposed yet. To allow for learning on fully dynamic graph streams, i.e., graphs that can change in their structure (addition/deletion of nodes/edge) and in their node/edge attributes, we propose a Marked Neural Spatio-Temporal Point Process (MNSTPP). It leverages a Dynamic Graph Neural Network to learn a Marked TPP that handles attributes and spatial data to model and predict any event in a graph stream.

Experimental evaluation is crucial in AI research, especially for assessing algorithms across diverse tasks. Many studies often evaluate a limited set of algorithms, failing to fully understand their strengths and weaknesses within a comprehensive portfolio. This paper introduces an Item Response Theory (IRT) based analysis tool for algorithm portfolio evaluation called AIRT-Module. Traditionally used in educational psychometrics, IRT models test question difficulty and student ability using responses to test questions. Adapting IRT to algorithm evaluation, the AIRT-Module contains a Shiny web application and the R package airt. AIRT-Module uses algorithm performance measures to compute anomalousness, consistency, and difficulty limits for an algorithm and the difficulty of test instances. The strengths and weaknesses of algorithms are visualised using the difficulty spectrum of the test instances. AIRT-Module offers a detailed understanding of algorithm capabilities across varied test instances, thus enhancing comprehensive AI method assessment. It is available at https://sevvandi.shinyapps.io/AIRT/ .

Causality is vital for understanding true cause-and-effect relationships between variables within predictive models, rather than relying on mere correlations, making it highly relevant in the field of Explainable AI. In an automated decision-making scenario, causal inference methods can analyze the underlying data-generation process, enabling explanations of a model's decision by manipulating features and creating counterfactual examples. These counterfactuals explore hypothetical scenarios where a minimal number of factors are altered, providing end-users with valuable information on how to change their situation. However, interpreting a set of multiple counterfactuals can be challenging for end-users who are not used to analyzing raw data records. In our work, we propose a novel multi-step pipeline that uses counterfactuals to generate natural language explanations of actions that will lead to a change in outcome in classifiers of tabular data using LLMs. This pipeline is designed to guide the LLM through smaller tasks that mimic human reasoning when explaining a decision based on counterfactual cases. We conducted various experiments using a public dataset and proposed a method of closed-loop evaluation to assess the coherence of the final explanation with the counterfactuals, as well as the quality of the content. Results are promising, although further experiments with other datasets and human evaluations should be carried out.

We introduce an information theoretic criterion for Bayesian network structure learning which we call quotient normalized maximum likelihood (qNML). In contrast to the closely related factorized normalized maximum likelihood criterion, qNML satisfies the property of score equivalence. It is also decomposable and completely free of adjustable hyperparameters. For practical computations, we identify a remarkably accurate approximation proposed earlier by Szpankowski and Weinberger. Experiments on both simulated and real data demonstrate that the new criterion leads to parsimonious models with good predictive accuracy.

Modern applications often leverage multiple views of a subject of study. Within neuroscience, there is growing interest in large-scale simultaneous recordings across multiple brain regions. Understanding the relationship between views (e.g., the neural activity in each region recorded) can reveal fundamental principles about the characteristics of each representation and about the system. However, existing methods to characterize such relationships either lack the expressivity required to capture complex nonlinearities, describe only sources of variance that are shared between views, or discard geometric information that is crucial to interpreting the data. Here, we develop a nonlinear neural network-based method that, given paired samples of high-dimensional views, disentangles low-dimensional shared and private latent variables underlying these views while preserving intrinsic data geometry. Across multiple simulated and real datasets, we demonstrate that our method outperforms competing methods. Using simulated populations of lateral geniculate nucleus (LGN) and V1 neurons we demonstrate our model's ability to discover interpretable shared and private structure across different noise conditions. On a dataset of unrotated and corresponding but randomly rotated MNIST digits, we recover private latents for the rotated view that encode rotation angle regardless of digit class, and places the angle representation on a 1-d manifold, while shared latents encode digit class but not rotation angle. Applying our method to simultaneous Neuropixels recordings of hippocampus and prefrontal cortex while mice run on a linear track, we discover a low-dimensional shared latent space that encodes the animal's position. We propose our approach as a general-purpose method for finding succinct and interpretable descriptions of paired data sets in terms of disentangled shared and private latent variables.

The emergence of deep learning models has revolutionized various industries over the last decade, leading to a surge in connected devices and infrastructures. However, these models can be tricked into making incorrect predictions with high confidence, leading to disastrous failures and security concerns. To this end, we explore the impact of adversarial attacks on multivariate time-series forecasting and investigate methods to counter them. Specifically, we employ untargeted white-box attacks, namely the Fast Gradient Sign Method (FGSM) and the Basic Iterative Method (BIM), to poison the inputs to the training process, effectively misleading the model. We also illustrate the subtle modifications to the inputs after the attack, which makes detecting the attack using the naked eye quite difficult. Having demonstrated the feasibility of these attacks, we develop robust models through adversarial training and model hardening. We are among the first to showcase the transferability of these attacks and defenses by extrapolating our work from the benchmark electricity data to a larger, 10-year real-world data used for predicting the time-to-failure of hard disks. Our experimental results confirm that the attacks and defenses achieve the desired security thresholds, leading to a 72.41% and 94.81% decrease in RMSE for the electricity and hard disk datasets respectively after implementing the adversarial defenses.

Current artificial intelligence (AI) models often focus on enhancing performance through meticulous parameter tuning and optimization techniques. However, the fundamental design principles behind these models receive comparatively less attention, which can limit our understanding of their potential and constraints. This comprehensive review explores the diverse design inspirations that have shaped modern AI models, i.e., brain-inspired artificial intelligence (BIAI). We present a classification framework that categorizes BIAI approaches into physical structure-inspired and human behavior-inspired models. We also examine the real-world applications where different BIAI models excel, highlighting their practical benefits and deployment challenges. By delving into these areas, we provide new insights and propose future research directions to drive innovation and address current gaps in the field. This review offers researchers and practitioners a comprehensive overview of the BIAI landscape, helping them harness its potential and expedite advancements in AI development.

Supervised learning-based software vulnerability detectors often fall short due to the inadequate availability of labelled training data. In contrast, Large Language Models (LLMs) such as GPT-4, are not trained on labelled data, but when prompted to detect vulnerabilities, LLM prediction accuracy is only marginally better than random guessing. In this paper, we explore a different approach by reframing vulnerability detection as one of anomaly detection. Since the vast majority of code does not contain vulnerabilities and LLMs are trained on massive amounts of such code, vulnerable code can be viewed as an anomaly from the LLM's predicted code distribution, freeing the model from the need for labelled data to provide a learnable representation of vulnerable code. Leveraging this perspective, we demonstrate that LLMs trained for code generation exhibit a significant gap in prediction accuracy when prompted to reconstruct vulnerable versus non-vulnerable code. Using this insight, we implement ANVIL, a detector that identifies software vulnerabilities at line-level granularity. Our experiments explore the discriminating power of different anomaly scoring methods, as well as the sensitivity of ANVIL to context size. We also study the effectiveness of ANVIL on various LLM families, and conduct leakage experiments on vulnerabilities that were discovered after the knowledge cutoff of our evaluated LLMs. On a collection of vulnerabilities from the Magma benchmark, ANVIL outperforms state-of-the-art line-level vulnerability detectors, LineVul and LineVD, which have been trained with labelled data, despite ANVIL having never been trained with labelled vulnerabilities. Specifically, our approach achieves $1.62\times$ to $2.18\times$ better Top-5 accuracies and $1.02\times$ to $1.29\times$ times better ROC scores on line-level vulnerability detection tasks.