Back to search

Aristotle University Of Thessaloniki (AUTH)

Machine learning from weakly-labelled and sparsely-positive data

Abstract

dc:description

The thesis deals with the topic of learning from weakly-labelled data that are also sparsely-positive. In a binary classification scenario, a weak label refers to a label that is used to describe a set of data points (or bag of instances), rather than a specific data point. On the other hand, the term ”sparsely-positive” implies that the salient characteristics that differentiate the positive from the negative class, are sporadically present both within the bag of instances and within the scope of each positive instance. This sparsity can be detrimental to learning algorithms, due to the low prevalence of the important characteristics and their overshadowing by other, unrelated to the task, characteristics. To deal with such problems, we first formalize them as a scenario of identifying rare positive events in large collections of signals. Then, we introduce a methodology for converting the original problem, to a form that can be solved efficiently by means of Multiple-Instance Learning (MIL). After transforming it to a suitable format, we propose using a MIL algorithm that is based on attention, in order to better handle the sparsely-positive nature of the problem. Finally, we build on this approach and introduce a multi-modal generalization of it, that is able to learn from multiple, weakly-labelled collections of signals, corresponding to data from multiple simultaneous sources (modalities or sensors), and again described by a single label. We continue by applying the analysis methodology to 2 real-world applications. First, we tackle the problem of detecting hand tremor caused by Parkinson’s disease (PD), using acceleration data recorded unobtrusively and in-the-wild via smartphone-embedded sensors. Furthermore, we apply the methodology to a sparsely-positive image classification problem from the medical domain: the detection of Diabetic Retinopathy (DR) in retinal images. Next, we demonstrate an application of the multi-modal approach to the problem of predicting if a person has PD by analyzing smartphone sensor data collected in-the-wild from 2 sources: hand acceleration data captured from the IMU sensor and typing data captured from the virtual keyboard of their smartphone. In the second part of the thesis, we introduce a novel semi-supervised approach for utilizing completely unlabelled bags in MIL problems. First, we show how to adapt the principle of Virtual Adversarial Training (VAT) to the MIL setting. VAT is a state-of-the-art approach for regular Semi-Supervised Learning (SSL) that encourages the model to be consistent across certain perturbations (referred to as virtual adversarial) of its inputs. Then, we go on to propose 2 additional variants of the main approach, which impose additional constraints on the form of the virtual adversarial perturbation. Finally, we apply our novel semi-supervised method to the PD tremor detection problem from the first part of the thesis, but now taking into account a large cohort of 454 subjects with acceleration data of the exact same nature as the initial cohort of 45 subjects, but lacking any form of tremor ground-truth.

Degree

thesis:*
Grantor dc:publisher
Aristotle University Of Thessaloniki (AUTH)
Year dc:date
2022

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Papadopoulos, Alexandros
  • Παπαδόπουλος, Αλέξανδρος

Subjects

dc:subject × 14

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/52892

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Papadopoulos, Alexandros; Παπαδόπουλος, Αλέξανδρος. Machine learning from weakly-labelled and sparsely-positive data. Aristotle University Of Thessaloniki (AUTH), 2022. http://hdl.handle.net/10442/hedi/52892