Search

Representation learning aims to distill useful knowledge from raw data and apply this knowledge to a wide range of applications. This ability to extract information that is useful not only for selected tasks but also generalizes to new settings is a key step towards artificial intelligence. In this thesis, we focus on representations derived through a specific type of generative model, i.e. variational autoencoders (VAEs). VAEs have several desirable properties. Thanks to the use of variational inference and the convenient model assumptions of Gaussian posteriors and a simple prior, VAEs are often easy to train and exhibit fast convergence. The probabilistic modelling formulation allows VAEs to derive a smooth latent representation of the raw data (i.e. semantically similar data samples are likely to be projected to nearby regions in the latent space). VAEs compress the raw data to a much lower dimension latent space. Working with the low-dimensional representations rather than the raw data can significantly reduces costs in memory and computation. With these advantages, VAEs have been widely applied to many applications, including robotics, drug discovery and digital content creation. Despite the widespread application of VAEs, improving the generative modelling of VAEs further remains an active research topic. In this thesis, we focus on two challenges in the VAE training: 1) over-regularized posterior distributions are often encountered in VAEs with Gaussian decoders and simple prior models; 2) the auto-encoding function may cause severe information drift and alter the information in the raw data in successive encodings. We propose solutions to both phenomena. Specifically, we optimize a variance parameter in the Gaussian decoder to balance competing loss terms in the ELBO objective. We adopt a flexible prior model that is implemented as a VAE in the latent space to mitigate the over-regularization effects. To reduce the information drift, we propose to modify the ELBO objective with a consistency loss that penalizes such drift. We show that these proposals can effectively address the challenges identified previously and improve the likelihood score of the VAEs. In addition to the contributions related to improving VAEs, we also demonstrate the power of VAEs' representation learning in two important machine learning applications. Firstly, we show that a VAE's ability to compress complicated, high-dimensional data is key to achieving good performance in anomaly detection. We design a VAE-LSTM anomaly detection system that can accurately identify anomalous effects in a time serious. Secondly, we show that a classifier which incorporates a VAE module can give better calibrated predictions. This is the result of a VAEs' capability of expressing the uncertainty between similar data samples in the spread of the posterior distribution and of identifying out-of-distribution samples.

Recent advances in miniaturized electronics, microfluidics, and electrochemical sensing platforms have paved the path for the realization of wearable and mobile biosensing systems that can be deployed to longitudinally access the circulating molecular information. Leveraging their non-/minimally-invasive nature and a high sampling rate, these systems provides people with a unprecedented accessibility to their health-related information, opening up new possibilities such as personalized medicine, precision nutrition, and performance enhancement. Towards realizing the envisioned biomonitoring ecosystem, various technical challenges remain unresolved, centering on the sampling and sensing of circulating molecular information. For example, sweat is widely targeted for non-invasive biomonitoring. However, current sweat sampling approaches are mainly stimulation-based (e.g., heat, exercise, iontophoresis), which is associated with user intervention and/or discomfort. Moreover, human subject studies are lacking to establish the physiological significance of the sweat biomarker levels. From the sensing side, electrochemical sensors are widely leveraged to quantify the analyte concentration levels due to their capability to provide sample-to-answer readouts. To accurately analyze untreated biofluids, novel sensor design and multi-dimension optimization are required to simultaneously achieve the required figure of merits, including sensitivity, selectivity, biofouling resistance, and generalizability. In my Ph.D. training, I devised sensor- and system-level solutions to address these challenges. As an immediate group of targets, my research mainly focusses on monitoring circulating drug levels as it can be leveraged to personalize dosing, evaluate adherence, and prevent abuse. In summary of my research work, the thesis is structured as follows: chapter 1 introduces the background of non-/minimally-invasive biomonitoring, including the existing bio-interfaces and electrochemical sensing mechanisms; chapter 2 describes the use of a thin hydrogel patch to simultaneously sample and analyze analyte concentrations in natural perspiration; chapter 3 further expands the innovation from chapter 2 and demonstrates a multi-modal human machine interface; chapter 4 describes an organohydrogel-based one-touch sensing interface for lithium therapy monitoring; chapter 5 focuses on resolving the challenges associated with voltammetry-based sensor—a simple recognition element-free sensing approach to target electroactive analyte—and demonstrates a wearable drug sensing smartwatch; chapter 6 presents my efforts to achieve a generalizable solution, where I demonstrated a microneedle-based aptamer sensor to target interstitial fluid analytes.