580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How can the architecture of fully connected layers be optimized for effective transfer learning in CNNs?
This theme explores methods to automatically learn and tune the structure and hyperparameters of fully connected (FC) layers within convolutional neural networks (CNNs) during transfer learning, addressing challenges of architectural design and overfitting particularly when adapting pre-trained models to target tasks with limited or domain-different data.
Key finding: Proposes AutoFCL, a Bayesian optimization framework to automatically learn the structure of FC layers tailored to the target dataset during transfer learning. Applied atop popular CNNs (VGG-16, ResNet, DenseNet, MobileNet,... Read more
Key finding: Introduces a dual transfer learning (DTL) framework that fine-tunes last layers, including FC layers, of pre-trained CNNs on both unlabeled and limited labeled medical images within the target domain. The approach counters... Read more
Key finding: Systematically evaluates transfer learning strategies focusing on fine-tuning FC layers pre-trained on large datasets like ImageNet, highlighting improved medical image classification despite limited target data. This work... Read more
2. What are the limitations of standard transfer learning models when applied to specialized domains such as 3D medical imaging, and how can custom models or advanced NAS (Neural Architecture Search) methods address these gaps?
This research area focuses on the challenges posed by using standard 2D transfer learning architectures pre-trained on natural images for complex, high-dimensional medical imaging tasks, especially volumetric data. It explores the performance differences between direct transfer, custom tailored 3D CNN architectures, and novel neural architecture transfer methods that adapt both topology and weights efficiently for various tasks with limited data.
Key finding: Demonstrates that standard 2D transfer learning models like InceptionV3, VGG16, and ResNet50 underperform compared to a specialized, custom 3D CNN tailored for 3D PET neuroimaging volumes, especially in multi-class... Read more
Key finding: Introduces Neural Architecture Transfer (NAT) to efficiently adapt pre-trained supernets with shared weights to new domains and multiple objectives via integrated online transfer learning and many-objective evolutionary... Read more
Key finding: Provides a comprehensive theoretical overview of transfer learning methodologies, outlining the differences between inductive, transductive, and unsupervised transfer learning. Emphasizes domain adaptation challenges where... Read more
3. How do transfer learning concepts integrate with domain-specific applications in engineering, signal processing, and computer graphics, and what are the methodological innovations enabling these applications?
This theme investigates the adaptation and integration of transfer learning methodologies beyond conventional image classification, focusing on application-driven innovations in geotechnical engineering, video coding, seismic imaging, face recognition, and 4D shape mapping. Key insights include techniques for dataset scarcity mitigation, architecture adaptation, and PDE-based mapping for complex task requirements.
Key finding: Examines transfer learning applications to tunnel support analysis in geotechnical engineering, demonstrating that pre-training neural networks on large datasets from one geological formation or a simplified analysis and... Read more
Key finding: Develops a feed-forward control system for optimal layered video transmission that dynamically adjusts layering parameters based on predicted macroblock complexity and network feedback. The controller optimizes perceptual... Read more
Key finding: Introduces a novel PDE-based methodology to compute implicit mappings between morphing surfaces of possibly different topologies by solving Laplace and transport equations in higher-dimensional embeddings. This method enables... Read more