In this dissertation, I carry out the research along the direction with particular focuses on scaling up the optimization of sparse learning for supervised and unsupervised learning problems. For the supervised learning, I firstly propose an asynchronous parallel solver to optimize the large-scale sparse learning model in a multithreading environment. Moreover, I propose a distributed framework to conduct the learning process when the dataset is distributed stored among different machines. Then the proposed model is further extended to the studies of risk genetic factors for Alzheimer's Disease (AD) among different research institutions, integrating a group feature selection framework to rank the top risk SNPs for AD. For the unsupervised learning problem, I propose a highly efficient solver, termed Stochastic Coordinate Coding (SCC), scaling up the optimization of dictionary learning and sparse coding problems. The common issue for the medical imaging research is that the longitudinal features of patients among different time points are beneficial to study together. To further improve the dictionary learning model, I propose a multi-task dictionary learning method, learning the different task simultaneously and utilizing shared and individual dictionary to encode both consistent and changing imaging features.
In this study, I aim to achieve multimodal brain image fusion by referring to some intrinsic properties of data, e.g. geometry of embedding structures where the commonly used image features reside. Since the image features investigated in this study share an identical embedding space, i.e. either defined on a brain surface or brain atlas, where a graph structure is easy to define, it is straightforward to consider the mathematically meaningful properties of the shared structures from the geometry perspective.
I first introduce the background of multimodal fusion of brain image data and insights of geometric properties playing a potential role to link different modalities. Then, several proposed computational frameworks either using the solid and efficient geometric algorithms or current geometric deep learning models are be fully discussed. I show how these designed frameworks deal with distinct geometric properties respectively, and their applications in the real healthcare scenarios, e.g. to enhanced detections of fetal brain diseases or abnormal brain development.
etc. Given a low resolution image, it aims to reconstruct a high resolution
image. The problem is ill-posed since there can be more than one high resolution
image corresponding to the same low-resolution image. To address this problem, a
number of machine learning-based approaches have been proposed.
In this dissertation, I present my works on single image super-resolution (SISR)
and accelerated magnetic resonance imaging (MRI) (a.k.a. super-resolution on MR
images), followed by the investigation on transfer learning for accelerated MRI reconstruction.
For the SISR, a dictionary-based approach and two reconstruction based
approaches are presented. To be precise, a convex dictionary learning (CDL)
algorithm is proposed by constraining the dictionary atoms to be formed by nonnegative
linear combination of the training data, which is a natural, desired property.
Also, two reconstruction-based single methods are presented, which make use
of (i)the joint regularization, where a group-residual-based regularization (GRR) and
a ridge-regression-based regularization (3R) are combined; (ii)the collaborative representation
and non-local self-similarity. After that, two deep learning approaches
are proposed, aiming at reconstructing high-quality images from accelerated MRI
acquisition. Residual Dense Block (RDB) and feedback connection are introduced
in the proposed models. In the last chapter, the feasibility of transfer learning for
accelerated MRI reconstruction is discussed.
To overcome the above data scarcity and generalization issues, in my dissertation, I first propose two unsupervised conventional machine learning algorithms, hyperbolic stochastic coding, and multi-resemble multi-target low-rank coding, to solve the incomplete data and missing label problem. I further introduce a deep multi-domain adaptation network to leverage the power of deep learning by transferring the rich knowledge from a large-amount labeled source dataset. I also invent a novel time-sequence dynamically hierarchical network that adaptively simplifies the network to cope with the scarce data.
To learn a large number of unseen concepts, lifelong machine learning enjoys many advantages, including abstracting knowledge from prior learning and using the experience to help future learning, regardless of how much data is currently available. Incorporating this capability and making it versatile, I propose deep multi-task weight consolidation to accumulate knowledge continuously and significantly reduce data requirements in a variety of domains. Inspired by the recent breakthroughs in automatically learning suitable neural network architectures (AutoML), I develop a nonexpansive AutoML framework to train an online model without the abundance of labeled data. This work automatically expands the network to increase model capability when necessary, then compresses the model to maintain the model efficiency.
In my current ongoing work, I propose an alternative method of supervised learning that does not require direct labels. This could utilize various supervision from an image/object as a target value for supervising the target tasks without labels, and it turns out to be surprisingly effective. The proposed method only requires few-shot labeled data to train, and can self-supervised learn the information it needs and generalize to datasets not seen during training.
A globally integrated carbon observation and analysis system is needed to improve the fundamental understanding of the global carbon cycle, to improve our ability to project future changes, and to verify the effectiveness of policies aiming to reduce greenhouse gas emissions and increase carbon sequestration. Building an integrated carbon observation system requires transformational advances from the existing sparse, exploratory framework towards a dense, robust, and sustained system in all components: anthropogenic emissions, the atmosphere, the ocean, and the terrestrial biosphere. The paper is addressed to scientists, policymakers, and funding agencies who need to have a global picture of the current state of the (diverse) carbon observations.
We identify the current state of carbon observations, and the needs and notional requirements for a global integrated carbon observation system that can be built in the next decade. A key conclusion is the substantial expansion of the ground-based observation networks required to reach the high spatial resolution for CO2 and CH4 fluxes, and for carbon stocks for addressing policy-relevant objectives, and attributing flux changes to underlying processes in each region. In order to establish flux and stock diagnostics over areas such as the southern oceans, tropical forests, and the Arctic, in situ observations will have to be complemented with remote-sensing measurements. Remote sensing offers the advantage of dense spatial coverage and frequent revisit. A key challenge is to bring remote-sensing measurements to a level of long-term consistency and accuracy so that they can be efficiently combined in models to reduce uncertainties, in synergy with ground-based data.
Bringing tight observational constraints on fossil fuel and land use change emissions will be the biggest challenge for deployment of a policy-relevant integrated carbon observation system. This will require in situ and remotely sensed data at much higher resolution and density than currently achieved for natural fluxes, although over a small land area (cities, industrial sites, power plants), as well as the inclusion of fossil fuel CO2 proxy measurements such as radiocarbon in CO2 and carbon-fuel combustion tracers. Additionally, a policy-relevant carbon monitoring system should also provide mechanisms for reconciling regional top-down (atmosphere-based) and bottom-up (surface-based) flux estimates across the range of spatial and temporal scales relevant to mitigation policies. In addition, uncertainties for each observation data-stream should be assessed. The success of the system will rely on long-term commitments to monitoring, on improved international collaboration to fill gaps in the current observations, on sustained efforts to improve access to the different data streams and make databases interoperable, and on the calibration of each component of the system to agreed-upon international scales.
Deposits of dark material appear on Vesta’s surface as features of relatively low-albedo in the visible wavelength range of Dawn’s camera and spectrometer. Mixed with the regolith and partially excavated by younger impacts, the material is exposed as individual layered outcrops in crater walls or ejecta patches, having been uncovered and broken up by the impact. Dark fans on crater walls and dark deposits on crater floors are the result of gravity-driven mass wasting triggered by steep slopes and impact seismicity. The fact that dark material is mixed with impact ejecta indicates that it has been processed together with the ejected material. Some small craters display continuous dark ejecta similar to lunar dark-halo impact craters, indicating that the impact excavated the material from beneath a higher-albedo surface. The asymmetric distribution of dark material in impact craters and ejecta suggests non-continuous distribution in the local subsurface. Some positive-relief dark edifices appear to be impact-sculpted hills with dark material distributed over the hill slopes.
Dark features inside and outside of craters are in some places arranged as linear outcrops along scarps or as dark streaks perpendicular to the local topography. The spectral characteristics of the dark material resemble that of Vesta’s regolith. Dark material is distributed unevenly across Vesta’s surface with clusters of all types of dark material exposures. On a local scale, some craters expose or are associated with dark material, while others in the immediate vicinity do not show evidence for dark material. While the variety of surface exposures of dark material and their different geological correlations with surface features, as well as their uneven distribution, indicate a globally inhomogeneous distribution in the subsurface, the dark material seems to be correlated with the rim and ejecta of the older Veneneia south polar basin structure. The origin of the dark material is still being debated, however, the geological analysis suggests that it is exogenic, from carbon-rich low-velocity impactors, rather than endogenic, from freshly exposed mafic material or melt, exposed or created by impacts.
The relation between flux and fluctuation is fundamental to complex physical systems that support and transport flows. A recently obtained law predicts monotonous enhancement of fluctuation as the average flux is increased, which in principle is valid but only for large systems. For realistic complex systems of small sizes, this law breaks down when both the average flux and fluctuation become large. Here we demonstrate the failure of this law in small systems using real data and model complex networked systems, derive analytically a modified flux-fluctuation law, and validate it through computations of a large number of complex networked systems. Our law is more general in that its predictions agree with numerics and it reduces naturally to the previous law in the limit of large system size, leading to new insights into the flow dynamics in small-size complex systems with significant implications for the statistical and scaling behaviors of small systems, a topic of great recent interest.
The apolipoprotein E (APOE) e4 allele is the most prevalent genetic risk factor for Alzheimer's disease (AD). Hippocampal volumes are generally smaller in AD patients carrying the e4 allele compared to e4 noncarriers. Here we examined the effect of APOE e4 on hippocampal morphometry in a large imaging database—the Alzheimer's Disease Neuroimaging Initiative (ADNI). We automatically segmented and constructed hippocampal surfaces from the baseline MR images of 725 subjects with known APOE genotype information including 167 with AD, 354 with mild cognitive impairment (MCI), and 204 normal controls. High-order correspondences between hippocampal surfaces were enforced across subjects with a novel inverse consistent surface fluid registration method. Multivariate statistics consisting of multivariate tensor-based morphometry (mTBM) and radial distance were computed for surface deformation analysis. Using Hotelling's T2 test, we found significant morphological deformation in APOE e4 carriers relative to noncarriers in the entire cohort as well as in the nondemented (pooled MCI and control) subjects, affecting the left hippocampus more than the right, and this effect was more pronounced in e4 homozygotes than heterozygotes. Our findings are consistent with previous studies that showed e4 carriers exhibit accelerated hippocampal atrophy; we extend these findings to a novel measure of hippocampal morphometry. Hippocampal morphometry has significant potential as an imaging biomarker of early stage AD.
Mild Cognitive Impairment (MCI) is a transitional stage between normal aging and dementia and people with MCI are at high risk of progression to dementia. MCI is attracting increasing attention, as it offers an opportunity to target the disease process during an early symptomatic stage. Structural magnetic resonance imaging (MRI) measures have been the mainstay of Alzheimer's disease (AD) imaging research, however, ventricular morphometry analysis remains challenging because of its complicated topological structure. Here we describe a novel ventricular morphometry system based on the hyperbolic Ricci flow method and tensor-based morphometry (TBM) statistics. Unlike prior ventricular surface parameterization methods, hyperbolic conformal parameterization is angle-preserving and does not have any singularities. Our system generates a one-to-one diffeomorphic mapping between ventricular surfaces with consistent boundary matching conditions. The TBM statistics encode a great deal of surface deformation information that could be inaccessible or overlooked by other methods. We applied our system to the baseline MRI scans of a set of MCI subjects from the Alzheimer's Disease Neuroimaging Initiative (ADNI: 71 MCI converters vs. 62 MCI stable). Although the combined ventricular area and volume features did not differ between the two groups, our fine-grained surface analysis revealed significant differences in the ventricular regions close to the temporal lobe and posterior cingulate, structures that are affected early in AD. Significant correlations were also detected between ventricular morphometry, neuropsychological measures, and a previously described imaging index based on fluorodeoxyglucose positron emission tomography (FDG-PET) scans. This novel ventricular morphometry method may offer a new and more sensitive approach to study preclinical and early symptomatic stage AD.