Insufficient training data poses significant challenges to training a deep convolutional neural network (CNN) to solve a target task. One common solution to this problem is to use transfer learning with pre-trained networks to apply knowledge learned from one domain with sufficient data to a new domain with limited data and avoid training a deep network from scratch. However, for such methods to work in a transfer learning setting, learned features from the source domain need to be generalizable to the target domain, which is not guaranteed since the feature space and distributions of the source and target data may be different. This thesis aims to explore and understand the use of orthogonal convolutional neural networks to improve learning of diverse, generic features that are transferable to a novel task. In this thesis, orthogonal regularization is used to pre-train deep CNNs to investigate if and how orthogonal convolution may improve feature extraction in transfer learning. Experiments using two limited medical image datasets in this thesis suggests that orthogonal regularization improves generality and reduces redundancy of learned features more effectively in certain deep networks for transfer learning. The results on feature selection and classification demonstrate the improvement in transferred features helps select more expressive features that improves generalization performance. To understand the effectiveness of orthogonal regularization on different architectures, this work studies the effects of residual learning on orthogonal convolution. Specifically, this work examines the presence of residual connections and its effects on feature similarities and show residual learning blocks help orthogonal convolution better preserve feature diversity across convolutional layers of a network and alleviate the increase in feature similarities caused by depth, demonstrating the importance of residual learning in making orthogonal convolution more effective.

Beta-Amyloid(Aβ) plaques and tau protein tangles in the brain are now widely recognized as the defining hallmarks of Alzheimer’s disease (AD), followed by structural atrophy detectable on brain magnetic resonance imaging (MRI) scans. However, current methods to detect Aβ/tau pathology are either invasive (lumbar puncture) or quite costly and not widely available (positron emission tomography (PET)). And one of the particular neurodegenerative regions is the hippocampus to which the influence of Aβ/tau on has been one of the research projects focuses in the AD pathophysiological progress. In this dissertation, I proposed three novel machine learning and statistical models to examine subtle aspects of the hippocampal morphometry from MRI that are associated with Aβ /tau burden in the brain, measured using PET images. The first model is a novel unsupervised feature reduction model to generate a low-dimensional representation of hippocampal morphometry for each individual subject, which has superior performance in predicting Aβ/tau burden in the brain. The second one is an efficient federated group lasso model to identify the hippocampal subregions where atrophy is strongly associated with abnormal Aβ/Tau. The last one is a federated model for imaging genetics, which can identify genetic and transcriptomic influences on hippocampal morphometry. Finally, I stated the results of these three models that have been published or submitted to peer-reviewed conferences and journals.

Alzheimer's disease (AD) is a neurodegenerative disease that damages the cognitive abilities of a patient. It is critical to diagnose AD early to begin treatment as soon as possible which can be done through biomarkers. One such biomarker is the beta-amyloid (Aβ) peptide which can be quantified using the centiloid (CL) scale. For identifying the Aβ biomarker, A deep learning model that can model AD progression by predicting the CL value for brain magnetic resonance images (MRIs) is proposed. Brain MRI images can be obtained through the Alzheimer's Disease Neuroimaging Initiative (ADNI) and Open Access Series of Imaging Studies (OASIS) datasets, however a single model cannot perform well on both datasets at once. Thus, A regularization-based continuous learning framework to perform domain adaptation on the previous model is also proposed which captures the latent information about the relationship between Aβ and AD progression within both datasets.

Communicating with computers through thought has been a remarkable achievement in recent years. This was made possible by the use of Electroencephalography (EEG). Brain-computer interface (BCI) relies heavily on Electroencephalography (EEG) signals for communication between humans and computers. With the advent ofdeep learning, many studies recently applied these techniques to EEG data to perform various tasks like emotion recognition, motor imagery classification, sleep analysis, and many more. Despite the rise of interest in EEG signal classification, very few studies have explored the MindBigData dataset, which collects EEG signals recorded at the stimulus of seeing a digit and thinking about it. This dataset takes us closer to realizing the idea of mind-reading or communication via thought. Thus classifying these signals into the respective digit that the user thinks about is a challenging task. This serves as a motivation to study this dataset and apply existing deep learning techniques to study it. Given the recent success of transformer architecture in different domains like Computer Vision and Natural language processing, this thesis studies transformer architecture for EEG signal classification. Also, it explores other deep learning techniques for the same. As a result, the proposed classification pipeline achieves comparable performance with the existing methods.

There is intense interest in adopting computer-aided diagnosis (CAD) systems, particularly those developed based on deep learning algorithms, for applications in a number of medical specialties. However, success of these CAD systems relies heavily on large annotated datasets; otherwise, deep learning often results in algorithms that perform poorly and lack generalizability. Therefore, this dissertation seeks to address this critical problem: How to develop efficient and effective deep learning algorithms for medical applications where large annotated datasets are unavailable. In doing so, we have outlined three specific aims: (1) acquiring necessary annotations efficiently from human experts; (2) utilizing existing annotations effectively from advanced architecture; and (3) extracting generic knowledge directly from unannotated images. Our extensive experiments indicate that, with a small part of the dataset annotated, the developed deep learning methods can match, or even outperform those that require annotating the entire dataset. The last part of this dissertation presents the importance and application of imaging in healthcare, elaborating on how the developed techniques can impact several key facets of the CAD system for detecting pulmonary embolism. Further research is necessary to determine the feasibility of applying these advanced deep learning technologies in clinical practice, particularly when annotation is limited. Progress in this area has the potential to enable deep learning algorithms to generalize to real clinical data and eventually allow CAD systems to be employed in clinical medicine at the point of care.

This work solves the problem of incorrect rotations while using handheld devices.Two new methods which improve upon previous works are explored. The first method
uses an infrared camera to capture and detect the user’s face position and orient the
display accordingly. The second method utilizes gyroscopic and accelerometer data
as input to a machine learning model to classify correct and incorrect rotations.
Experiments show that these new methods achieve an overall success rate of 67%
for the first and 92% for the second which reaches a new high for this performance
category. The paper also discusses logistical and legal reasons for implementing this
feature into an end-user product from a business perspective. Lastly, the monetary
incentive behind a feature like irRotate in a consumer device and explore related
patents is discussed.

In recent years, Convolutional Neural Networks (CNNs) have been widely used in not only the computer vision community but also within the medical imaging community. Specifically, the use of pre-trained CNNs on large-scale datasets (e.g., ImageNet) via transfer learning for a variety of medical imaging applications, has become the de facto standard within both communities.

However, to fit the current paradigm, 3D imaging tasks have to be reformulated and solved in 2D, losing rich 3D contextual information. Moreover, pre-trained models on natural images never see any biomedical images and do not have knowledge about anatomical structures present in medical images. To overcome the above limitations, this thesis proposes an image out-painting self-supervised proxy task to develop pre-trained models directly from medical images without utilizing systematic annotations. The idea is to randomly mask an image and train the model to predict the missing region. It is demonstrated that by predicting missing anatomical structures when seeing only parts of the image, the model will learn generic representation yielding better performance on various medical imaging applications via transfer learning.

The extensive experiments demonstrate that the proposed proxy task outperforms training from scratch in six out of seven medical imaging applications covering 2D and 3D classification and segmentation. Moreover, image out-painting proxy task offers competitive performance to state-of-the-art models pre-trained on ImageNet and other self-supervised baselines such as in-painting. Owing to its outstanding performance, out-painting is utilized as one of the self-supervised proxy tasks to provide generic 3D pre-trained models for medical image analysis.

Recently, a well-designed and well-trained neural network can yield state-of-the-art results across many domains, including data mining, computer vision, and medical image analysis. But progress has been limited for tasks where labels are difficult or impossible to obtain. This reliance on exhaustive labeling is a critical limitation in the rapid deployment of neural networks. Besides, the current research scales poorly to a large number of unseen concepts and is passively spoon-fed with data and supervision.

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.

The ubiquity of single camera systems in society has made improving monocular depth estimation a topic of increasing interest in the broader computer vision community. Inspired by recent work in sparse-to-dense depth estimation, this thesis focuses on sparse patterns generated from feature detection based algorithms as opposed to regular grid sparse patterns used by previous work. This work focuses on using these feature-based sparse patterns to generate additional depth information by interpolating regions between clusters of samples that are in close proximity to each other. These interpolated sparse depths are used to enforce additional constraints on the network’s predictions. In addition to the improved depth prediction performance observed from incorporating the sparse sample information in the network compared to pure RGB-based methods, the experiments show that actively retraining a network on a small number of samples that deviate most from the interpolated sparse depths leads to better depth prediction overall.

This thesis also introduces a new metric, titled Edge, to quantify model performance in regions of an image that show the highest change in ground truth depth values along either the x-axis or the y-axis. Existing metrics in depth estimation like Root Mean Square Error(RMSE) and Mean Absolute Error(MAE) quantify model performance across the entire image and don’t focus on specific regions of an image that are hard to predict. To this end, the proposed Edge metric focuses specifically on these hard to classify regions. The experiments also show that using the Edge metric as a small addition to existing loss functions like L1 loss in current state-of-the-art methods leads to vastly improved performance in these hard to classify regions, while also improving performance across the board in every other metric.

Unsupervised learning of time series data, also known as temporal clustering, is a challenging problem in machine learning. This thesis presents a novel algorithm, Deep Temporal Clustering (DTC), to naturally integrate dimensionality reduction and temporal clustering into a single end-to-end learning framework, fully unsupervised. The algorithm utilizes an autoencoder for temporal dimensionality reduction and a novel temporal clustering layer for cluster assignment. Then it jointly optimizes the clustering objective and the dimensionality reduction objective. Based on requirement and application, the temporal clustering layer can be customized with any temporal similarity metric. Several similarity metrics and state-of-the-art algorithms are considered and compared. To gain insight into temporal features that the network has learned for its clustering, a visualization method is applied that generates a region of interest heatmap for the time series. The viability of the algorithm is demonstrated using time series data from diverse domains, ranging from earthquakes to spacecraft sensor data. In each case, the proposed algorithm outperforms traditional methods. The superior performance is attributed to the fully integrated temporal dimensionality reduction and clustering criterion.