Neural networks (NNs) can achieved high performance in various fields such as computer vision, and natural language processing. However, deploying NNs in resource-constrained safety-critical systems has challenges due to uncertainty in the prediction caused by out-of-distribution data, and hardware non-idealities. To address the challenges of deploying NNs in resource-constrained safety-critical systems, this paper summarizes the (4th year) PhD thesis work that explores scalable and efficient methods for uncertainty estimation and reduction in deep learning, with a focus on Computation-in-Memory (CIM) using emerging resistive non-volatile memories. We tackle the inherent uncertainties arising from out-of-distribution inputs and hardware non-idealities, crucial in maintaining functional safety in automated decision-making systems. Our approach encompasses problem-aware training algorithms, novel NN topologies, and hardware co-design solutions, including dropout-based \emph{binary} Bayesian Neural Networks leveraging spintronic devices and variational inference techniques. These innovations significantly enhance OOD data detection, inference accuracy, and energy efficiency, thereby contributing to the reliability and robustness of NN implementations.

In open-domain Question Answering (QA), dense retrieval is crucial for finding relevant passages for answer generation. Typically, contrastive learning is used to train a retrieval model that maps passages and queries to the same semantic space. The objective is to make similar ones closer and dissimilar ones further apart. However, training such a system is challenging due to the false negative issue, where relevant passages may be missed during data annotation. Hard negative sampling, which is commonly used to improve contrastive learning, can introduce more noise in training. This is because hard negatives are those closer to a given query, and thus more likely to be false negatives. To address this issue, we propose a novel contrastive confidence regularizer for Noise Contrastive Estimation (NCE) loss, a commonly used loss for dense retrieval. Our analysis shows that the regularizer helps dense retrieval models be more robust against false negatives with a theoretical guarantee. Additionally, we propose a model-agnostic method to filter out noisy negative passages in the dataset, improving any downstream dense retrieval models. Through experiments on three datasets, we demonstrate that our method achieves better retrieval performance in comparison to existing state-of-the-art dense retrieval systems.

The advent of instruction-tuned language models that convincingly mimic human writing poses a significant risk of abuse. For example, such models could be used for plagiarism, disinformation, spam, or phishing. However, such abuse may be counteracted with the ability to detect whether a piece of text was composed by a language model rather than a human. Some previous approaches to this problem have relied on supervised methods trained on corpora of confirmed human and machine-written documents. Unfortunately, model under-specification poses an unavoidable challenge for neural network-based detectors, making them brittle in the face of data shifts, such as the release of further language models producing still more fluent text than the models used to train the detectors. Other previous approaches require access to the models that may have generated a document in question at inference or detection time, which is often impractical. In light of these challenges, we pursue a fundamentally different approach not relying on samples from language models of concern at training time. Instead, we propose to leverage representations of writing style estimated from human-authored text. Indeed, we find that features effective at distinguishing among human authors are also effective at distinguishing human from machine authors, including state of the art large language models like Llama 2, ChatGPT, and GPT-4. Furthermore, given a handful of examples composed by each of several specific language models of interest, our approach affords the ability to predict which model generated a given document.

Emergencies and critical incidents often unfold rapidly, necessitating a swift and effective response. In this research, we introduce a novel approach to identify and classify emergency situations from social media posts and direct emergency messages using an open source Large Language Model, LLAMA2. The goal is to harness the power of natural language processing and machine learning to assist public safety telecommunicators and huge crowds during countrywide emergencies. Our research focuses on developing a language model that can understand users describe their situation in the 911 call, enabling LLAMA2 to analyze the content and offer relevant instructions to the telecommunicator, while also creating workflows to notify government agencies with the caller's information when necessary. Another benefit this language model provides is its ability to assist people during a significant emergency incident when the 911 system is overwhelmed, by assisting the users with simple instructions and informing authorities with their location and emergency information.

In the last decade, the research of artificial intelligence on medical images has attracted researchers' interest. One of the most popular directions is automated medical image segmentation (MIS) using deep learning, which aims to automatically assign labels to pixels so that the pixels with the same label from a segmented object. However, in the past years a strong trend of highlighting or cherry-picking improper metrics to show particularly high scores close to 100% was revealed in scientific publishing of MIS studies [1]. In addition, even though there are some papers that evaluate image segmentation results from different perspectives, the implementation of their evaluation algorithms is inconsistent. This is due to the lack of a universal metric library in Python for standardized and reproducible evaluation. Therefore, we propose to develop an open-source publicly available Python package seg-metrics, which aims to evaluate the performance of MIS models.

This research investigates the transferability of Automatic Speech Recognition (ASR)-robust Natural Language Understanding (NLU) models from controlled experimental conditions to practical, real-world applications. Focused on smart home automation commands in Urdu, the study assesses model performance under diverse noise profiles, linguistic variations, and ASR error scenarios. Leveraging the UrduBERT model, the research employs a systematic methodology involving real-world data collection, cross-validation, transfer learning, noise variation studies, and domain adaptation. Evaluation metrics encompass task-specific accuracy, latency, user satisfaction, and robustness to ASR errors. The findings contribute insights into the challenges and adaptability of ASR-robust NLU models in transcending controlled environments.

The RecSys Challenge 2023, presented by ShareChat, consists to predict if an user will install an application on his smartphone after having seen advertising impressions in ShareChat & Moj apps. This paper presents the solution of 'Team UMONS' to this challenge, giving accurate results (our best score is 6.622686) with a relatively small model that can be easily implemented in different production configurations. Our solution scales well when increasing the dataset size and can be used with datasets containing missing values.

In this paper, we introduce SeizNet, a closed-loop system for predicting epileptic seizures through the use of Deep Learning (DL) method and implantable sensor networks. While pharmacological treatment is effective for some epilepsy patients (with ~65M people affected worldwide), one out of three suffer from drug-resistant epilepsy. To alleviate the impact of seizure, predictive systems have been developed that can notify such patients of an impending seizure, allowing them to take precautionary measures. SeizNet leverages DL techniques and combines data from multiple recordings, specifically intracranial electroencephalogram (iEEG) and electrocardiogram (ECG) sensors, that can significantly improve the specificity of seizure prediction while preserving very high levels of sensitivity. SeizNet DL algorithms are designed for efficient real-time execution at the edge, minimizing data privacy concerns, data transmission overhead, and power inefficiencies associated with cloud-based solutions. Our results indicate that SeizNet outperforms traditional single-modality and non-personalized prediction systems in all metrics, achieving up to 99% accuracy in predicting seizure, offering a promising new avenue in refractory epilepsy treatment.

Feature selection in noisy label scenarios remains an understudied topic. We propose a novel genetic algorithm-based approach, the Noise-Aware Multi-Objective Feature Selection Genetic Algorithm (NMFS-GA), for selecting optimal feature subsets in binary classification with noisy labels. NMFS-GA offers a unified framework for selecting feature subsets that are both accurate and interpretable. We evaluate NMFS-GA on synthetic datasets with label noise, a Breast Cancer dataset enriched with noisy features, and a real-world ADNI dataset for dementia conversion prediction. Our results indicate that NMFS-GA can effectively select feature subsets that improve the accuracy and interpretability of binary classifiers in scenarios with noisy labels.

Large-scale administrative data is a common input in early warning systems for college dropout in higher education. Still, the terminology and methodology vary significantly across existing studies, and the implications of different modeling decisions are not fully understood. This study provides a systematic evaluation of contributing factors and predictive performance of machine learning models over time and across different student groups. Drawing on twelve years of administrative data at a large public university in the US, we find that dropout prediction at the end of the second year has a 20% higher AUC than at the time of enrollment in a Random Forest model. Also, most predictive factors at the time of enrollment, including demographics and high school performance, are quickly superseded in predictive importance by college performance and in later stages by enrollment behavior. Regarding variability across student groups, college GPA has more predictive value for students from traditionally disadvantaged backgrounds than their peers. These results can help researchers and administrators understand the comparative value of different data sources when building early warning systems and optimizing decisions under specific policy goals.