Browsing by Subject "Machine Learning"

Virtually every consumer product available on the market today contains some form of fossil fuel-based polymer. However, these materials pose environmental, human health, and economic concerns due to their enduring presence in the global ecosystem and their degradation products. Addressing this crisis necessitates scalable production of biodegradable alternatives, such as polyhydroxyalkanoates (PHAs). PHAs are presented as promising substitutes due to their biodegradability, biocompatibility, and the potential for complete renewable utilization post-degradation, but a current challenge to widespread use of these materials lies in understanding the quantitative relationship between the structural characteristics of PHAs, their environmental interactions, and their degradation rates to enhance their industrial production and distribution. To bridge this knowledge gap, the dissertation outlines a comprehensive approach involving the development of a specialized dataset, the application of machine learning (ML) models to predict degradation rates based on structural and environmental factors, and the experimental validation of these predictions. The first part of this research focuses on assembling a manually curated dataset from the extensive, available open-access literature, aimed at understanding the effects of structural and environmental features on PHA degradation. The second part leverages this dataset through ML modeling, employing techniques like random forest regression to predict degradation profiles with over 80% accuracy. This methodology enables a deeper understanding of the complex interplay between chemical structures and degradation properties, surpassing traditional trial-and-error approaches. The final part of this research aims to complete an iterative workflow for dataset development by validating ML model predictions through physical experiments, enriching the original dataset with comprehensive experimental data on PHA degradation in hydrolytic environments with contact angle, molecular weight, and thermal property characterizations. The incorporation of experimental findings into the ML dataset, particularly through expanded ML techniques that emphasize pairwise feature importance such as explainable boosting machines (EBM), helps in pinpointing critical factors influencing PHA degradation, such as environmental temperature and material properties. The model performances indicate a strong performance of manually assembled literature-based datasets when predicting degradation rate for PHAs. In conclusion, a data science-based framework has been developed for exploring PHA biopolyester degradation and explores the combination of features of the material and its environment that integrates the structure, properties, and experimentally verified degradation profiles of the material. This workflow will be a useful and generalizable pipeline for PHAs and other polymers to expand the biopolymer design space with degradation in mind.

Deep learning has demonstrated remarkable success in developing vision intelligence across a variety of application domains, including autonomous driving, facial recognition, medical image analysis, \etc.However, developing such vision systems poses significant challenges, particularly in relation to ensuring efficiency, interpretability, and fairness. Efficiency requires a model to leverage the least possible computational resources while preserving performance relative to more computationally-demanding alternatives, which is essential for the practical deployment of large-scale models in real-time applications. Interpretability demands a model to align with the domain-specific knowledge of the task it addresses while having the capability for case-based reasoning. This characteristic is especially crucial in high-stakes areas such as healthcare, criminal justice, and financial investment. Fairness ensures that computer vision models do not perpetuate or exacerbate societal biases in downstream applications such as web image search, text-guided image generation, \etc. In this dissertation, I will discuss the contributions that I have made in advancing vision intelligence regarding to efficiency, interpretability and fairness in computer vision models.

The first part of this dissertation will focus on how to design computer vision models to efficiently process very large images.We propose a novel CNN architecture termed { \em Zoom-In Network} that leverages a hierarchical attention sampling mechanisms to select important regions of images to process. Such approach without processing the entire image yields outstanding memory efficiency while maintaining classification accuracy on various tiny object image classification datasets.

The second part of this dissertation will discuss how to build post-hoc interpretation method for deep learning models to obtain insights reasoned from the predictions.We propose a novel image and text insight-generation framework based on attributions from deep neural nets. We test our approach on an industrial dataset and demonstrate our method outperforms competing methods.

Finally, we study fairness in large vision-language models.More specifically, we examined gender and racial bias in text-based image retrieval for neutral text queries. In an attempt to address bias in the test-time phase, we proposed post-hoc bias mitigation to actively balance the demographic group in the image search results. Experiments on multiple datasets show that our method can significantly reduce bias while maintaining satisfactory retrieval accuracy at the same time.

My research in enhancing vision intelligence via developments in efficiency, interpretability, and fairness, has undergone rigorous validation using publicly available benchmarks and has been recognized at leading peer-reviewed machine learning conferences.This dissertation has sparked interest within the AI community, emphasizing the importance of improving computer vision models through these three critical dimensions, namely, efficiency, interpretability and fairness.

Objective:To develop a machine learning model that integrates deep learning image features and radiomics features to classify lymph nodes as benign or malignant in Non-Small Cell Lung Cancer (NSCLC). Methods: The dataset comprises contrast-enhanced CT scans from 541 lung cancer patients before surgery, collected at a Shanghai Hospital between July 2015 and December 2017 under an IRB study. It includes 1,237 lymph nodes, identified from preoperative CT scans due to enlargement and confirmed as non-small cell lung cancer (NSCLC) via surgical pathology. Lymph node classification into malignant or benign categories utilized in postoperative pathological reports. Our method employs a dual radiomic feature extraction strategy. The deep image features (DIF) were derived from the final convolutional layer of a pre-trained VGG-16 encoder network to characterize the lymph node’s image texture. A total of nine 2D shape-based radiomic features (RF) are extracted based on the Py-radiomics calculation toolbox to characterize lymph node morphological information. And ninety-two handcrafted radiomic features (HRF) are extracted. The extracted DIF, RF, and HRF were combined and fed into a Random Forest classifier for the benign and malignant lymph node classification. The random forest classifier was trained following an 8:2 train/test split ratio and evaluated using Area Under the Curve (AUC), Receiver Operating Characteristic (ROC), and p-value, and 5-fold cross-validation was also employed to objectively evaluate model performance.

Results: The mean AUC for the Random Forest classifier using only 2D shape features is 0.691, while mean AUC for the classifier employing only DIF is 726. Utilizing both DIF and HRF for classification resulted in an average AUC of 0.724, whereas integrating RF with DIF achieved superior classification efficacy, boasting the highest average AUC of 0.746. All results were considered statistically significant with a p-value of less than 0.05. Conclusion: The combination of image texture analysis refers to DIF with morphological information offers an enhanced characterization ability to classify lymph nodes as benign or malignant from CT images for lung NSCLC patients.

Background

Methods

Results

Conclusions

Objective

Methods

Results

Conclusions

Background

Methods

Results

Conclusions

Trial registration

Background

Methods

Results

Conclusion

The advent of Artificial Intelligence (AI) propels the real world into a new era characterized by remarkable design innovations and groundbreaking design automation, primarily fueled by Deep Neural Networks (DNN). At the heart of this transformation is the progress in Automated Machine Learning (AutoML), notably Neural Architecture Search (NAS). NAS lays a robust foundation for developing algorithms capable of automating design processes to determine the optimal architecture for academic benchmarks. However, the real challenge emerges when adapting NAS for Applied Machine Learning (AML) scenarios: navigating the complex terrain of design space exploration and exploitation. This complexity arises due to the heterogeneity of data and architectures required by real-world AML problems, an aspect that traditional NAS approaches struggle to address fully.

To bridge this gap, our research emphasizes creating a flexible search space that reduces reliance on human-derived architectural assumptions. We introduce innovative techniques aimed at refining search algorithms to accommodate greater flexibility. By carefully examining and enhancing search spaces and methodologies, we empower NAS solutions to cater to practical AML problems. This enables the exploration of broader search spaces, better performance potential, and lower search process costs.

We start by challenging homogeneous search space design for multi-modality 3D representations, proposing ``PIDS'' to enable joint dimension and interaction search for 3D point cloud segmentation. We consider two axes on adapting point cloud operators toward multi-modality data with density, geometry, and order varieties, achieving significant mIOU improvement on segmentation benchmarks over the state-of-the-art 3D models.To implement our approach efficiently in recommendation systems, we develop ``NASRec'' to support heterogeneous building operators and propose practical solutions to improve the quality of NAS on Click-Through Rates (CTR) prediction. We propose an end-to-end full architecture search with minimal human priors. We provide practical solutions to tackle scalability and heterogeneity challenges in NAS, outperforming manually designed models and existing NAS models on various CTR benchmarks. Finally, we pioneer our effort on industry-scale CTR benchmarks and propose DistDNAS to optimize search and serving efficiency, producing smaller and better recommendation models on a large-scale CTR benchmark. Intuited by the discoveries in NAS, we additionally uncover the underlying theoretical foundations of residual learning on computer vision foundation research and envision the prospects of our research on Artificial Intelligence, including Large Language Models, Generative AI, and beyond.

A significant gap in education lies in the need for mechanisms that enable early detection of potentially at-risk students. Through access to an earlier prediction of student performance, instructors are given ample time to meet with and assist under-achieving students. As with any prediction modeling problem, there are many predictors to choose from when formulating a model. Previous related works have shown limited success in predicting course performance using students' personal and socioeconomic traits. Students learn by asking clarifying questions. Therefore, discussion boards have been a staple of learning at the university level for years.

This research aims to utilize participation in discussion forums to predict final student performance. Using students' course grades at roughly the halfway point in the term and various discussion forum predictors, our model predicts the students' final percentage score. Using the model's prediction, instructors can speak with at-risk students and discuss ways to improve. The student grades and discussion board participation datasets are gathered from graduate-level Electrical and Computer Engineering (ECE) courses at Duke University. Various classical machine learning models are explored, with random forest yielding the highest accuracy. This random forest model, trained on discussion forum participation data, surpasses other similarly trained state-of-the-art models.

Furthermore, related research attempts the classification problem of predicting what discrete letter grade a student will earn. This is not an accurate representation of a student's performance, and therefore, we attempt the regression problem of predicting the exact percentage a student will earn. A significant finding of this research is that our random forest model can predict student performance with an average error of approximately 2.3%. Additionally, our random forest model can generalize to a different graduate-level course and make performance predictions with an average error of 3.3%.

The final important finding is that a model including discussion board predictors outperforms another whose sole predictor is the students' halfway point grade. This indicates that discussion forums hold significant value in determining final performance. We envision that the knowledge from our findings and our optimal random forest model can enable instructors to identify and support potentially at-risk students preemptively.

Purpose: To develop a novel radiomic quantification framework to quantify the impact of image resampling and discretization on radiomic texture characterization performance.

Methods: The study employed 251 CT scans of a Credence Cartridge phantom (consisting of 10 texture materials) with different image acquisition parameters. Each material was segmented using a pre-defined cylindrical mask. Different image pre-processing workflows including 5 resampling methods (no resampling, trilinear, and nearest resampling to both 1mm³ and 5mm³) and 8 discretization methods (fixed bin size of 25,50,75,100 and fixed bin counts of 8,16,32,64) were randomly applied. 75 radiomic texture features (including 24GLCM-based, 16GLRLM-based, 16GLSZM-based, 14GLDM-based, and 5NGTDM-based) were extracted from each material to characterize its textural attributes. Three machine learning models including logistic regression (LR), random forest (RF), and supporting vector machine (SVM) were developed to identify 10 materials based on the extracted features, and grid search was adopted to optimize the model hyperparameters. The model performance was evaluated on 10-class macro-AUC with 5-fold cross-validation.

Results: Three models successfully classified 10 materials with macro-AUC=0.9941±0.0081, 0.9979±0.0040, and 0.9957±0.0067 for LR, RF, and SVM, respectively. Across 8 different discretization methods, an increasing trend in performance can be observed when the original CT was discretized to a larger gray level range: performance improved by 0.0038 with bin sizes decreasing from 100-25, and by 0.0074 with bin counts increasing from 8 to 64. Among 5 resampling methods, resampling CT to an isotropic voxel spacing showed an improved prediction performance (0.9942±0.0075/0.9944±0.0073 for trilinear/nearest resampling to 1mm³ and 5mm³, respectively) over no interpolation (0.9862±0.0228), with minimal performance discrepancies observed among two different interpolation algorithms. In addition, no statistically significant differences were observed across five folds.

Conclusion: The proposed framework successfully quantified the dependence of radiomics texture characterization on image resampling and discretization.