Browsing by Author "Patel, Ankit"

Now showing 1 - 7 of 7
  • Thumbnail Image
    ItemEmbargo
    A Computational Analysis of Meal Events Using Food Diaries and Continuous Glucose Monitors
    (2023-04-21) Pai, Amruta; Sabharwal, Ashutosh; Allen, Genevera; Patel, Ankit; Beier, Margaret; Kerr, David
    Diet self-management, through its effect on weight and glycemic control, is one of the cornerstones of Type 2 Diabetes (T2D) prevention and management. A quantitative understanding of bio-behavioral mechanisms of diet is needed to create effective diet self-management tools. Smartphone diet-tracking applications and continuous glucose monitors (CGMs) are emerging devices that enable dense sampling of an individual's diet. Research in diet analysis of app-based food diaries and CGMs have mainly focused on developing aggregate measures of nutrient intake and glucose responses. However, innovative computational analysis is required to infer actionable insights. In this thesis, we develop computational measures for various bio-behavioral aspects of diet by leveraging meal event data collected with food diaries and CGMs. First, we establish recurrent consumption measures across meal events to characterize habitual behavior in an individual's diet. We leverage a large publicly available MyFitnessPal (MFP) food diary dataset to provide novel insights on differences in habitual behavior across individuals and temporal contexts. Next, we develop calorie compensation measures to characterize self-regulatory behavior. A quantitative analysis of calorie compensation measures on the MFP dataset reveals significant meal compensation patterns and their impact on adherence to self-set calorie goals. Finally, we designed an observational study using the MFP app and CGMs to evaluate the impact of meal events on glycemic control in adults with varying hemoglobin a1c levels. We developed elevated meal event count to characterize mealtime glucose responses by exploiting its association with hemoglobin a1c. Elevated meal event count significantly affected glycemic control, suggesting its value as a novel event-driven glycemic target metric. This thesis highlights the value of using CGMs and food diaries to broaden our understanding of diet. The developed measures augment existing intake measures and could be used as a digital bio-behavioral markers to personalize diet self-management strategies.
  • Thumbnail Image
    Item
    Computational chromatography: A machine learning strategy for demixing individual chemical components in complex mixtures
    (PNAS, 2022) Bajomo, Mary M.; Ju, Yilong; Zhou, Jingyi; Elefterescu, Simina; Farr, Corbin; Zhao, Yiping; Neumann, Oara; Nordlander, Peter; Patel, Ankit; Halas, Naomi J.; Laboratory for Nanophotonics
    Surface-enhanced Raman spectroscopy (SERS) holds exceptional promise as a streamlined chemical detection strategy for biological and environmental contaminants compared with current laboratory methods. Priority pollutants such as polycyclic aromatic hydrocarbons (PAHs), detectable in water and soil worldwide and known to induce multiple adverse health effects upon human exposure, are typically found in multicomponent mixtures. By combining the molecular fingerprinting capabilities of SERS with the signal separation and detection capabilities of machine learning (ML), we examine whether individual PAHs can be identified through an analysis of the SERS spectra of multicomponent PAH mixtures. We have developed an unsupervised ML method we call Characteristic Peak Extraction, a dimensionality reduction algorithm that extracts characteristic SERS peaks based on counts of detected peaks of the mixture. By analyzing the SERS spectra of two-component and four-component PAH mixtures where the concentration ratios of the various components vary, this algorithm is able to extract the spectra of each unknown component in the mixture of unknowns, which is then subsequently identified against a SERS spectral library of PAHs. Combining the molecular fingerprinting capabilities of SERS with the signal separation and detection capabilities of ML, this effort is a step toward the computational demixing of unknown chemical components occurring in complex multicomponent mixtures.
  • Thumbnail Image
    Item
    Efficient Machine Vision Using Computational Cameras
    (2016-10-21) Chen, George; Veeraraghavan, Ashok; Patel, Ankit
    Computational cameras, powered by novel optics and advanced signal processing algorithms, has emerged as a powerful imaging tool that brings orders of magni- tude performance improvements over current camera technology. However, existing computer vision pipelines are still built around conventional digital cameras. In this thesis, we propose a novel computer vision framework that integrates computational cameras for machine vision applications. I explore two possible ways of improving the energy-efficiency and cost-effectiveness under such proposed framework. We first introduce ASP Vision, a jointly designed sensor + deep learning system for visual recognition tasks. ASP Vision utilizes angle sensitive pixels (ASP) to optically compute the first layer of convolutional neural networks (CNN), resulting 10x savings in sensing energy and bandwidth, and 2-4% savings in CNN FLOPs, while achieving similar performance compared to traditional deep learning pipelines. We then present FPA-CS, a focal plane array based compressive sensing architecture that provides a 15x cost savings in high-resolution shortwave infrared (SWIR) video acquisition.
  • Thumbnail Image
    Item
    Interpreting Deep Neural Networks and Beyond: Visualization, Learning Dynamics, and Disentanglement
    (2021-01-15) Nie, Weili; Patel, Ankit
    Despite their great success, deep neural networks are always considered black boxes. In various domains such as self-driving cars and tumor diagnosis, it is crucial to know the reasoning behind a decision made by a neural network. In addition, a better understanding of current deep learning methods will inspire the development of more principled computational approaches with better robustness and interpretability. In this thesis, we focus on interpreting and improving deep neural networks from the following three perspectives: visualization, learning dynamics, and disentanglement. First, the demand for human interpretable explanations for model decisions has driven the development of visualization techniques. In particular, a class of backpropagation-based visualizations has attracted much attention recently, including Saliency Map, DeconvNet, and Guided Backprop (GBP). However, we find that there exist some perplexing behaviors: DeconvNet and GBP are more human-interpretable but less class-sensitive than Saliency Map. Motivated by this, we develop a theory that shows that GBP and DeconvNet are essentially doing image reconstruction, which is unrelated to network decisions. This analysis, together with various experiments, implies that those human-interpretable visualizations do not always reveal the inner working mechanisms of deep neural networks. Second, although generative adversarial networks (GANs) have been one of the most powerful deep generative models, they are notoriously difficult to train and the reasons underlying their (non-)convergence behaviors are still not completely understood. To this end, we conduct a non-asymptotic analysis of local convergence in GAN training dynamics by evaluating the eigenvalues of its Jacobian near the equilibrium. The analysis reveals that to ensure a good convergence rate, two factors should be avoided: (i) Phase Factor, i.e., the Jacobian has complex eigenvalues with a large imaginary-to-real ratio, and (ii) Conditioning Factor, i.e., the Jacobian is ill-conditioned. Thus, we propose a new regularization method called JARE that addresses both factors by construction. Third, disentanglement learning aims to make representations in neural networks more disentangled and human interpretable. However, we find that current disentanglement methods have several limitations: 1) difficulty with high-resolution images, 2) neglecting the existence of a trade-off between learning disentangled representations and controllable generation, and 3) non-identifiability due to the unsupervised setting. To overcome these limitations, we propose new losses and network architectures based on StyleGAN [karras et al., 2019] for semi-supervised high-resolution disentanglement learning. Experimental results show that using very limited supervision significantly improves disentanglement quality and that the proposed method can generalize well to unseen images in the tasks of semantic fine-grained image editing. Looking forward, with more efforts and meaningful interactions in these three directions, we believe that we can improve our understanding of both the successes and failures of current deep learning methods, and develop more robust, interpretable and flexible AI systems.
  • Thumbnail Image
    Item
    Mapping the (un)known: Spatially-driven Frameworks for characterization of Disease Heterogeneity
    (2023-08-10) Krishnan, Santhoshi Navaneetha; Patel, Ankit; Rao, Arvind
    In recent years, novel treatment approaches such as immunotherapy, where we boost an individual's immune system for better cancer targeting, are being increasingly trialed and adopted for various cancer pathologies. However, immune system function response is highly variable, especially in cancers such as Pancreatic Ductal Adenocarcinoma(PDAC), with many patients either not responding or suffering from adverse effects. Recent work has highlighted the importance of spatial quantification of tumor and immune cell subtypes to provide insights into treatment prognosis. My proposed work involves leveraging machine learning and spatial statistical methods to build frameworks that better characterize the (un)known tumor microenvironment, utilizing the inherent spatial structure extracted from imaging data. We validate these frameworks using commonly accessible histopathological imaging data such as hematoxylin and eosin(H&E)-stained whole slide images and multiplexed immunofluorescence(mIF) images. For this purpose, I explore three distinct methodological approaches to quantify tumor-immune interactions in a patient cohort consisting of 6 different pancreatic diseases, including PDAC. In the first approach, I utilize the concepts of geographically weighted regression (GWR) and a density function-based classification model to compute a spatially salient construction that can discriminate the six classes of lesions. This framework, GaWRDenMap, showed significant discriminant ability in multiple pairwise comparisons compared to abundance-based metrics, like the Morisita-Horn index. Our framework was able to highly discriminate between PDAC and non-cancerous chronic pancreatitis(CP), which is challenging for pathologists to discern based on the arrangement and structure of the cells under the microscope. Next, I repurpose an existing discriminative feature-based dictionary learning method to identify sub-regions of the tumor microenvironment representative of the disease. Initial results on mIF images from CP and PDAC patients from the same pancreatic cohort point to the excellent discriminant capabilities of the model while providing a visually interpretable dictionary output representative of the given cohort. Finally, I propose a novel cell-graph-based Cell-Graph ATtention (CGAT) network for precisely classifying PDAC and its precursor lesions only utilizing available spatial and phenotypic information as input features. The self-attention mechanism facilitates the identification of tissue regions and novel cell-cell interaction patterns characteristic of the disease. The proposed methods move away from single-number summaries by providing a more comprehensive representation of the state of the microenvironment. This thesis lays the groundwork for adopting these tools into not just exploratory biological research but also as diagnostic tools in the clinical setting.
  • Thumbnail Image
    ItemEmbargo
    Objective Sociability and Impulsivity Measures for Dimensional Assessment of Mental Health
    (2024-04-16) Lamichhane, Bishal; Sabharwal, Ashutosh; Sano, Akane; Patel, Ankit; Beier, Margaret
    Mental health disorders have a high prevalence and are on the rise globally. Subjective self-reports used for mental health assessment, with their recall and reporting biases, pose a scalability challenge for accurate and frequent assessment. Additionally, the existing nosology of disorders has rampant comorbidity and heterogeneity, creating challenges in the diagnosis and treatment. Instead of the dichotomous diagnostic labeling of mental health based on subjective self-reports, as is currently practiced, assessing mental health along the objectively measured bio-behavioral dimensions of functioning could provide a foundation for scalable, robust, and accurate mental healthcare. Towards the goal of improved mental health assessment, this thesis investigated objective measures of sociability and impulsivity, the bio-behavioral dimensions of functioning implicated in several mental health disorders. Sociability represents an individual’s tendency to interact and socialize with others. Sociability deficits and altered sociability patterns are commonly implicated across several mental health disorders. In this thesis, we propose using free-living audio to obtain sociability measures objectively. We developed a deep learning-based audio processing pipeline, ECoNet, that addresses the challenges of free-living audio processing to infer sociability measures accurately. The inferred sociability measures were significantly correlated with dimensional mental health measures in a clinical study comprising participants with diverse mental health conditions. Free-living audio also complemented the conventional mobile sensing modalities for dimensional mental health score prediction. Our results establish the potential of free-living audio-based objective sociability measures to represent mental health state. Impulsivity represents an individual’s tendency to act on urges without sufficient forethought about one’s action. Several mental health disorders commonly implicate heightened impulsivity. In this thesis, we propose using multimodal behavioral, physiological, and neurobiological measurements to model impulsivity objectively. We conducted a clinical study with mood disorder participants and discovered the complementarity of objective modalities to model impulsivity and predict the clinical outcome of suicidality with high accuracy. The complementarity of modalities was additionally validated with a large open-source dataset. To exploit the richness of neurobiological measurements for predicting impulsivity, we developed ImpulsNet. ImpulsNet is a lightweight deep learning model based on multi-task learning that outperforms existing models in the literature to predict impulsivity. Our results establish the feasibility of objectively modeling impulsivity for mental health applications. This thesis provides a foundation for dimensional mental health assessment using objective sociability and impulsivity measures. Further validation in larger studies should be pursued in future work. Other objectively assessed dimensions of functioning should also be explored to better represent mental health in the vastness of bio-behavioral space.
  • Thumbnail Image
    Item
    Representing Formal Languages: A Comparison Between Finite Automata and Recurrent Neural Networks
    (2019-04-19) Michalenko, Joshua James; Patel, Ankit; Baraniuk , Richard
    We investigate the internal representations that a recurrent neural network (RNN) uses while learning to recognize a regular formal language. Specially, we train a RNN on positive and negative examples from a regular language, and ask if there is a simple decoding function that maps states of this RNN to states of the minimal deterministic fnite automaton (MDFA) for the language. Our experiments show that such a decoding function indeed exists, and that it maps states of the RNN not to MDFA states, but to states of an abstraction obtained by clustering small sets of MDFA states into “superstates”. A framework for performing large scale systematic representation analysis between the two language models is discussed. Quantitative analysis surprisingly shows that linear decoding functions are suffcient for the task and an analysis of a range of abstraction functions is given. A qualitative analysis reveals new interpretations of how RNNs implement hierarchical priors during the language recognition task. Overall, the results suggest a strong structural relationship between internal representations used by RNNs and fnite automata, and explain the well-known ability of RNNs to recognize formal grammatical structure.