H-ensemble: An Information Theoretic Approach to Reliable Few-Shot Multi-Source-Free Transfer. (2024)

A framework that learns an optimal linear combination of pre-trained source models (a.k.a. source ensemble) to a given target task by maximizing transferability. Leading to a multi-source transfer learning model that does not reply on having source model details or source training data. (more info)

Efficient Prediction of Model Transferability in Semantic Segmentation Tasks. (2023)

Efficiently adapts transferability metrics for semantic segmentation. Based on the adapted metric, an transferability-weighted fine-tuning approach for transfering between semantic segmentation tasks is proposed. (more info)

Finding the most transferable task for MRI brain segmentation. (2022)

A knowledge and transferability-based source task selection framework for the transfer learning among MRI brain segmentation tasks. Empirical studies reveals that modality information and segmentation mask structural similarity can significantly improve the transferability estimation between different segmentation tasks. (more info)

OTCE: A transferability metric for cross-domain cross-task representations. (2021)

A new transferability metric is presented for transfer learning across heterogeneous data distributions (a.k.a. domains) and distinct tasks. It characterizes transferability as a combination of domain difference and task difference and explicitly evaluates them from data in a unified framework. (more info)

An information-theoretic metric to transferability for task transfer learning (2019)

Given a common input domain, a transferability metric estimates to what extent knowledge from a source task can help in learning a target task. We present a novel metric, rooted in information theory and statistics, to efficiently compute task transferability between classification problems. (more info)

Enhancing Continuous Domain Adaptation with Multi-Path Transfer Curriculum (2024)

A continuous domain adaptation model, W-MPOT, which adapts source domain training data to distant target domain with the help of intermediate domains. It finds an optimal ‘‘transfer curriculum’’ of intermediate domains and regularizes the optimal-transport based continous adaptation with novel multi-path constraints. (more info)

Class-conditioned domain generalization via wasserstein distributional robust optimization (2021)

Domain generalization aims at learning a domain-agnostic model from multiple source domains for any unseen target domain in the future. In this work, we solve the domain generalization problem using the concept of Wasserstein distributional robust optimization. (more info)

Few-Shot Cross Domain Battery Capacity Estimation (2021)

In data-driven battery capacity estimation, different battery sizes, testing environments and historical load patterns cause large domain shifts between training and testing data. We propose an optimal transport based domain adaptation method that is both efficient and accurate for few-shot cross-domain capacity estimation. (more info)

Joint PVL segmenation and hand function classification (2021)

A semi-supervised multitask learning framework to jointly learn PVL lesion segmentation and manual ability classification for MRI scans. Two clinically related auxiliary tasks are incorporated to improve the classification accuracy while requiring only a small amount of manual annotations. (more info)

Maximal correlation embedding network for multilabel learning with missing labels (2019)

Missing label is a common problem in real world multi-label learning datasets, where some of the labels associated with a sample may be missing due to sensor or human annotation error. We propose a robust maximal correlation embedding network for multi-label classification with missing labels. It has been successfully applied to multimedia and human context recognition data. (more info)

Multimodal emotion recognition: extracting public and private information (2019)

An end-to-end emotion recognition framework using both visual and audio input. It learns a multi-modal representation that captures both dependences between different input modalities, and modal-dependent information in each modality. (more info)

Learning Persistent Community Structures in Dynamic Networks via Topological Data Analysis. (2024)

A novel deep graph clustering framework with temporal consistency regularization on inter-community struc- tures, inspired by the concept of minimal network topolog- ical changes within short intervals. (more info)

Topology-Preserving Hard Pixel Mining for Tubular Structure Segmentation (2023)

A skeleton-based hard pixel mining strategy for training end-to-end segmentation networks for tubular structures, penalizing topology-relevant yet mis-segmented pixels in overall cost function. Skeletal hard pixels are mined through several simple logical and morphological operations which are fast and scalable. (more info)

Explainable Trajectory Representation through Dictionary Learning (2023)

A dictionary learning approach for encoding trajectories on a network in an explainable and efficient way. It learns a set of frequently traversed path segments (pathlets) that optimally reconstruct every trajectory by concatenation. The resulting representation is naturally sparse and encodes strong spatial semantics. (more info)

Urban mobility pattern mining based on regional dependencies (2018, 2019)

A novel mobility pattern mining algorithm that learns the dynamics between different spatial regions from taxi trip data. It proposes kernelized ACE, which efficiently finds latent embeddings for trip origins and destinations while maintaining the correlation between them. The resulting embedding can be used to extract strong mobility patterns while partitioning the city into non-overlapping regions at the same time. (more info)

Travel time prediction from limited GPS floating cars (2017, 2019)

A trajectory-based travel time prediction algorithm when only a small number of GPS floating cars are available. It learns the travel time patterns of a compact set of frequently shared paths from historical data. Given a travel time prediction query, we identify the current travel time pattern from recent trajectories, then infer its travel time in the near future. (more info)

Knowledge-based trajectory completion (2016)

A framework to "densify" sparse GPS trajectories without relying on the roadmap data. We are able to recover geometric details of complex junctions, and improve the accuracy of real life traffic trajectories. (more info)

Data-driven map matching (2013)

A robust multitrack map matching algorithm for sparse and noisy GPS trajectories. It simultaneously learns the regular structures in taxi trajectories and map match all trajectories simultaneously. (more info)

Mining leader-follower realtionships from GPS trajectories (2012)

A case study on understanding the leader-follower phenomenum in the group movement of cows from their trajectories. We use geometric and correlation based constraint to extract leader-follower relationships. (more info)

Session-based recommendation with temporal dynamics (2023)

A session-based recommendation framework for large volunteer networks that employs temporal dynamics to capture uncertainty caused by the changing structure of volunteers’ participation behaviour. (more info)

Optimising Self-organized Volunteer Behaviors during COVID-19 Pandemic (2022)

A quantitive study on how volunteers collaborate to achieve rapid mobilisation during the COVID-19 outbreak, using the concept of self-organisation. It proposes a data-driven framework to investigate when and how self-organisation emerged during the pandemic response and how it relates to effectiveness of volunteer organisations in general. (more info)

Last edited on November 2^nd 2024 10:12AM (Time Zone: EDT).
Powered by jemdoc + new design.