abstract: In many fields one needs to build predictive models for a set of related machine learning tasks, such as information retrieval, computer vision and biomedical informatics. Traditionally these tasks are treated independently and the inference is done separately for each task, which ignores important connections among the tasks. Multi-task learning aims at simultaneously building models for all tasks in order to improve the generalization performance, leveraging inherent relatedness of these tasks. In this thesis, I firstly propose a clustered multi-task learning (CMTL) formulation, which simultaneously learns task models and performs task clustering. I provide theoretical analysis to establish the equivalence between the CMTL formulation and the alternating structure optimization, which learns a shared low-dimensional hypothesis space for different tasks. Then I present two real-world biomedical informatics applications which can benefit from multi-task learning. In the first application, I study the disease progression problem and present multi-task learning formulations for disease progression. In the formulations, the prediction at each point is a regression task and multiple tasks at different time points are learned simultaneously, leveraging the temporal smoothness among the tasks. The proposed formulations have been tested extensively on predicting the progression of the Alzheimer's disease, and experimental results demonstrate the effectiveness of the proposed models. In the second application, I present a novel data-driven framework for densifying the electronic medical records (EMR) to overcome the sparsity problem in predictive modeling using EMR. The densification of each patient is a learning task, and the proposed algorithm simultaneously densify all patients. As such, the densification of one patient leverages useful information from other patients. / Dissertation/Thesis / Ph.D. Computer Science 2014

Multi-task learning (MTL) is a rapidly growing field in the world of autonomous vehicles, particularly in the area of computer vision. Autonomous vehicles are heavily reliant on computer vision technology for tasks such as object detection, object segmentation, and object tracking. The complexity of sensor data and the multiple tasks involved in autonomous driving can make it challenging to design effective systems. MTL addresses these challenges by training a single model to perform multiple tasks simultaneously, utilizing shared representations to learn common concepts between a group of related tasks, and improving data efficiency. In this thesis, we proposed a scalable MTL system for object detection that can be used to construct any MTL network with different scales and shapes. The proposed system is an extension to the state-of-art algorithm called Mask RCNN. It is designed to overcome the limitations of learning multiple objects in multi-label learning. To demonstrate the effectiveness of the proposed system, we built three different networks using it and evaluated their performance on the state-of-the-art BDD100k dataset. Our experimental results demonstrate that the proposed MTL networks outperform a base single-task network, Mask RCNN, in terms of mean average precision at 50 (mAP50). Specifically, the proposed MTL networks achieved a mAP50 of 66%, while the base network only achieved 53%. Furthermore, we also conducted comparisons between the proposed MTL networks to determine the most efficient way to group tasks together in order to create an optimal MTL network for object detection on the BDD100k dataset.

We consider a multi-task learning problem, where an agent is presented a number of N reinforcement learning tasks. To solve this problem, we are interested in studying the gradient approach, which iteratively updates an estimate of the optimal policy using the gradients of the value functions. The classic policy gradient method, however, may be expensive to implement in the multi-task settings as it requires access to the gradients of all the tasks at every iteration. To circumvent this issue, in this paper we propose to study an incremental policy gradient method, where the agent only uses the gradient of only one task at each iteration. Our main contribution is to provide theoretical results to characterize the performance of the proposed method. In particular, we show that incremental policy gradient methods converge to the optimal value of the multi-task reinforcement learning objectives at a sublinear rate O(1/√k), where k is the number of iterations. To illustrate its performance, we apply the proposed method to solve a simple multi-task variant of GridWorld problems, where an agent seeks to find an policy to navigate effectively in different environments. / Master of Science / First, we introduce a popular machine learning technique called Reinforcement Learning (RL), where an agent, such as a robot, uses a policy to choose an action, like moving forward, based on observations from sensors like cameras. The agent receives a reward that helps judge if the policy is good or bad. The objective of the agent is to find a policy that maximizes the cumulative reward it receives by repeating the above process. RL has many applications, including Cruise autonomous cars, Google industry automation, training ChatGPT language models, and Walmart inventory management. However, RL suffers from task sensitivity and requires a lot of training data. For example, if the task changes slightly, the agent needs to train the policy from the beginning. This motivates the technique called Multi-Task Reinforcement Learning (MTRL), where different tasks give different rewards and the agent maximizes the sum of cumulative rewards of all the tasks. We focus on the incremental setting where the agent can only access the tasks one by one randomly. In this case, we only need one agent and it is not required to know which task it is performing. We show that the incremental policy gradient methods we proposed converge to the optimal value of the MTRL objectives at a sublinear rate O(1/ √ k), where k is the number of iterations. To illustrate its performance, we apply the proposed method to solve a simple multi-task variant of GridWorld problems, where an agent seeks to find an policy to navigate effectively in different environments.

The Transformer model is a very recent, fast and powerful discovery in neural machine translation. We experiment with multi-task learning for enriching the source side of the Transformer with linguistic resources to provide it with additional information to learn linguistic and world knowledge better. We analyze two approaches: the basic shared model with multi-tasking through simple data manipulation, and multi-decoder models. We test joint models for machine translation (MT) and POS tagging, dependency parsing and named entity recognition as the secondary tasks. We evaluate them in comparison with the baseline and with dummy, linguistically unrelated tasks. We focus primarily on the standard- size data setting for German-to-Czech MT. Although our enriched models did not significantly outperform the baseline, we empirically document that (i) the MT models benefit from the secondary linguistic tasks; (ii) considering the amount of training data consumed, the multi-tasking models learn faster; (iii) in low-resource conditions, the multi-tasking significantly improves the model; (iv) the more fine-grained annotation of the source as the secondary task, the higher benefit to MT.

abstract: Multi-task learning (MTL) aims to improve the generalization performance (of the resulting classifiers) by learning multiple related tasks simultaneously. Specifically, MTL exploits the intrinsic task relatedness, based on which the informative domain knowledge from each task can be shared across multiple tasks and thus facilitate the individual task learning. It is particularly desirable to share the domain knowledge (among the tasks) when there are a number of related tasks but only limited training data is available for each task. Modeling the relationship of multiple tasks is critical to the generalization performance of the MTL algorithms. In this dissertation, I propose a series of MTL approaches which assume that multiple tasks are intrinsically related via a shared low-dimensional feature space. The proposed MTL approaches are developed to deal with different scenarios and settings; they are respectively formulated as mathematical optimization problems of minimizing the empirical loss regularized by different structures. For all proposed MTL formulations, I develop the associated optimization algorithms to find their globally optimal solution efficiently. I also conduct theoretical analysis for certain MTL approaches by deriving the globally optimal solution recovery condition and the performance bound. To demonstrate the practical performance, I apply the proposed MTL approaches on different real-world applications: (1) Automated annotation of the Drosophila gene expression pattern images; (2) Categorization of the Yahoo web pages. Our experimental results demonstrate the efficiency and effectiveness of the proposed algorithms. / Dissertation/Thesis / Ph.D. Computer Science 2011

Over 4 million avoidable hospital admissions result from medication errors (IMS Insitute for Healthcare Informatics, 2013). Human error accounts for 80% of all medical errors (Palmieri, DeLucia, Peterson, Ott, & Green, 2008). Medication administration is a complex process. It is important to understand the cognitive load (CL) of Registered Nurses (RNs) working in an electronic health record environment to identify the risk factors of medication errors. The purpose of this study is to investigate the factors that influence the CL of RNs during medication administration who are working in an electronic health record environment. Simulated medication administration scenarios with varying degrees of multi-tasking were completed with 30 participants. When RNs multi-task during medication administration their CL increases. Furthermore, RNs who have poor sleep quality cannot process high-level tasks as well as those RNs who report a good sleep quality. Future work can limit EEG lead placement to the frontal channels of the EEG. Furthermore, replication of this study with a larger sample and a broader range of competing tasks is indicated.

This thesis uses a multi-tasking long short-term memory (LSTM) model to investigate the correlation between containment strategies, climate change, and the number of COVID-19 transmissions and deaths. The study focuses on examining the accuracy of different factors in predicting the number of daily confirmed cases and deaths cases to further explore the correlation between different factors and cases. The initial assessment results suggest that containment strategies, specifically vaccination policies, have a more significant impact on the accuracy of predicting daily confirmed cases and deaths from COVID-19 compared to climate factors such as the daily average surface 2-meter temperature. Additionally, the study reveals that there are unpredictable effects on predictive accuracy resulting from the interactions among certain impact factors. However, the lack of interpretability of deep learning models poses a significant challenge for real-world applications. This study provides valuable insights into understanding the correlation between the number of daily confirmed cases, daily deaths, containment strategies, and climate change, and highlights areas for further research. It is important to note that while the study reveals a correlation, it does not imply causation, and further research is needed to understand the trends of the pandemic.

A smart manufacturing network connects machines via sensing, communication, and actuation networks. The data generated from the networks are used in data-driven modeling and decision-making to improve quality, productivity, and flexibility while reducing the cost. This dissertation focuses on improving the data-driven modeling of the quality-process relationship in smart manufacturing networks. The quality-process variable relationships are important to understand for guiding the quality improvement by optimizing the process variables. However, several challenges emerge. First, the big data sets generated from the manufacturing network may be information-poor for modeling, which may lead to high data transmission and computational loads and redundant data storage. Second, the data generated from connected machines often contain inexplicit similarities due to similar product designs and manufacturing processes. Modeling such inexplicit similarities remains challenging. Third, it is unclear how to select representative data sets for modeling in a manufacturing network setting, considering inexplicit similarities. In this dissertation, a data filtering method is proposed to select a relatively small and informative data subset. Multi-task learning is combined with latent variable decomposition to model multiple connected manufacturing processes that are similar-but-non-identical. A data filtering and modeling framework is also proposed to filter the manufacturing data for manufacturing network modeling adaptively. The proposed methodologies have been validated through simulation and the applications to real manufacturing case studies. / Doctor of Philosophy / The advancement of the Internet-of-Things (IoT) integrates manufacturing processes and equipment into a network. Practitioners analyze and apply the data generated from the network to model the manufacturing network to improve product quality. The data quality directly affects the modeling performance and decision effectiveness. However, the data quality is not well controlled in a manufacturing network setting. In this dissertation, we propose a data quality assurance method, referred to as data filtering. The proposed method selects a data subset from raw data collected from the manufacturing network. The proposed method reduces the complexity in modeling while supporting decision effectiveness. To model the data from multiple similar-but-non-identical manufacturing processes, we propose a latent variable decomposition-based multi-task learning model to study the relationships between the process variables and product quality variable. Lastly, to adaptively determine the appropriate data subset for modeling each process in the manufacturing network, we further proposed an integrated data filtering and modeling framework. The proposed integrated framework improved the modeling performance of data generated by babycare manufacturing and semiconductor manufacturing.

The integration of Fog-Cloud computing in manufacturing has given rise to a new paradigm called Fog manufacturing. Fog manufacturing is a form of distributed computing platform that integrates Fog-Cloud collaborative computing strategy to facilitate responsive, scalable, and reliable data analysis in manufacturing networks. The computation services provided by Fog-Cloud computing can effectively support quality prediction, process monitoring, and diagnosis efforts in a timely manner for manufacturing processes. However, the communication and computation resources for Fog-Cloud computing are limited in Fog manufacturing. Therefore, it is significant to effectively utilize the computation services based on the optimal computation task offloading, scheduling, and hardware autoscaling strategies to finish the computation tasks on time without compromising on the quality of the computation service. A prerequisite for adapting such optimal strategies is to accurately predict the run-time metrics (e.g., Time-latency) of the Fog nodes by capturing their inherent stochastic nature in real-time. It is because these run-time metrics are directly related to the performance of the computation service in Fog manufacturing. Specifically, since the computation flow and the data querying activities vary between the Fog nodes in practice. The run-time metrics that reflect the performance in the Fog nodes are heterogenous in nature and the performance cannot be effectively modeled through traditional predictive analysis. In this thesis, a multi-task learning methodology is adopted to predict the run-time metrics that reflect performance in Fog manufacturing by addressing the heterogeneities among the Fog nodes. A Fog manufacturing testbed is employed to evaluate the prediction accuracies of the proposed model and benchmark models. The proposed model can be further extended in computation tasks offloading and architecture optimization in Fog manufacturing to minimize the time-latency and improve the robustness of the system. / Master of Science / Smart manufacturing aims at utilizing Internet of things (IoT), data analytics, cloud computing, etc. to handle varying market demand without compromising the productivity or quality in a manufacturing plant. To support these efforts, Fog manufacturing has been identified as a suitable computing architecture to handle the surge of data generated from the IoT devices. In Fog manufacturing computational tasks are completed locally through the means of interconnected computing devices called Fog nodes. However, the communication and computation resources in Fog manufacturing are limited. Therefore, its effective utilization requires optimal strategies to schedule the computational tasks and assign the computational tasks to the Fog nodes. A prerequisite for adapting such strategies is to accurately predict the performance of the Fog nodes. In this thesis, a multi-task learning methodology is adopted to predict the performance in Fog manufacturing. Specifically, since the computation flow and the data querying activities vary between the Fog nodes in practice. The metrics that reflect the performance in the Fog nodes are heterogenous in nature and cannot be effectively modeled through conventional predictive analysis. A Fog manufacturing testbed is employed to evaluate the prediction accuracies of the proposed model and benchmark models. The results show that the multi-task learning model has better prediction accuracy than the benchmarks and that it can model the heterogeneities among the Fog nodes. The proposed model can further be incorporated in scheduling and assignment strategies to effectively utilize Fog manufacturing's computational services.

BACKGROUND. Multimodal Machine Learning is a powerful paradigm that capitalizes on the complementary predictive capabilities of different data modalities, such as text, image, time series. This approach allows for an extremely diverse feature space, which proves useful for combining different real-world tasks into a single model. Current architectures in the field of multimodal learning often integrate feature representations in parallel, a practice that not only limits their interpretability but also creates a reliance on the availability of specific modalities. Interpretability and robustness to missing inputs are particularly important in clinical decision support systems. To address these issues, the iGH Research Group at EPFL proposed a modular sequential input fusion called Modular Decision Support Network (MoDN). MoDN was tested on unimodal tabular inputs for multitask outputs and was shown to be superior to its monolithic parallel counterparts, while handling any number and combination of inputs and providing continuous real-time predictive feedback. AIM. We aim to extend MoDN to MultiModN with multimodal inputs and compare the benefits and limitations of sequential fusion with a state-of-the-art parallel fusion (P-Fusion) baseline.METHODS & FINDINGS. We align our experimental setup with a previously published P-Fusion baseline, focusing on two binary diagnostic predictive tasks (presence of pleural effusion and edema) in a popular multimodal clinical benchmark dataset (MIMIC).We perform four experiments: 1) comparing MultiModN to P-Fusion, 2) extending the architecture to multiple tasks, 3) exploring MultiModN's inherent interpretability in several metrics, and 4) testing its ability to be resistant to biased missingness by simulating missing not at random (MNAR) data during training and flipping the bias at inference. We show that MultiModN's sequential architecture does not compromise performance compared with the P-Fusion baseline, despite the added advantages of being multitask, composable and inherently interpretable. The final experiment shows that MultiModN resists catastrophic failure from MNAR data, which is particularly prevalent in clinical settings. / Multimodal maskininlärning är ett kraftfullt paradigm som utnyttjar de kompletterande prediktiva egenskaperna hos olika datamodaliteter, såsom text, bild, tidsserier. Detta tillvägagångssätt möjliggör ett extremt varierat funktionsutrymme, vilket visar sig vara användbart för att kombinera olika verkliga uppgifter i en enda modell. Nuvarande arkitekturer för multimodal inlärning integrerar ofta funktionsrepresentationer parallellt, en praxis som inte bara begränsar deras tolkningsbarhet utan också skapar ett beroende av tillgängligheten av specifika modaliteter. Tolkningsbarhet och robusthet mot saknade indata är särskilt viktigt i kliniska beslutsstödsystem. För att lösa dessa problem har forskargruppen iGH vid EPFL föreslagit en modulär sekventiell fusion av indata som kallas Modular Decision Support Network (MoDN). MoDN testades på unimodala tabulära indata för multitask-utdata och visade sig vara överlägsen sina monolitiska parallella motsvarigheter, samtidigt som den hanterar alla antal och kombinationer av indata och ger kontinuerlig prediktiv feedback i realtid. Vårt mål är att utöka MoDN till MultiModN med multimodala indata och jämföra fördelarna och begränsningarna med sekventiell fusion med en toppmodern baslinje för parallell fusion (P-Fusion). Vi anpassar vår experimentuppsättning till en tidigare publicerad P-Fusion-baslinje, med fokus på två binära diagnostiska prediktiva uppgifter (närvaro av pleural effusion och ödem) i en populär multimodal klinisk benchmark datauppsättning (MIMIC), som omfattar bilder, text, tabelldata och tidsserier. Vi utför fyra experiment och visar att MultiModN:s sekventiella arkitektur inte försämrar prestandan jämfört med P-Fusions baslinje, trots de extra fördelarna med att vara multitasking, komponerbar och tolkningsbar i sin egen rätt. Det sista experimentet visar att MultiModN motstår katastrofala fel från MNAR-data, vilket är särskilt vanligt i kliniska miljöer.