ProfIvor Tsang

Learning from Noisy Supervision / Ambiguity:
-Curriculum Loss: Robust Learning and Generalization against Label Corruption
-MAAN: Marginalized Average Attentional Network for Weakly-Supervised Learning
-Masking: A New Perspective of Noisy Supervision
-Stochastic Multi-Channel Ranking for Noisy BCI Signals
-Safeguarded Dynamic Label Regression for Noisy Supervision and Open-Set Classes
-Contrastive Additive Noise Network: Deep Learning from Noisy Labels With Quality Embedding
-Co-teaching+: Cross-Update from Disagreement
-Co-teaching: Robust Training of Deep Neural Networks with Extremely Noisy Labels
-CoarsenRank: Distribution Robust Rank Aggregration
-Weakly Labeled Support Vector Machines (WellSVMs) via Label Generation
-Co-labeling for multi-view weakly labeled learning
-Maximum Entropy Discrimination with Feature Concatenation (MEDFC) for learning with missing views and missing labels


Reinforcement Learning and Imitation Learning:
-GIRL: Generative Intrinsic Reward Learning using One-life Demonstration for Imitation Learning


Big Data Analytics and Scalable Machine Learning Algorithms:
-Core Vector Machines (CVMs) for Classification, Regression, Semi-supervised learning, Dimension reduction on Big Data
-Ball Vector Machines (BVMs) for Classification on Bigger Data
-Matching Pursuit LASSO (MPL) for Sparse Recovery over Big Dictionaries with Million Atoms and Sparse Recovery over Batch Signals
-Approximate Conditional Gradient Descent on Large-Scale Multi-Class Classification
-Riemannian Pursuit for Big Matrix Recovery
-Laplacian Embedded Support Vector Regression (LapESVR) for Scalable Manifold Regularization on a Big Graph


Extremely Many Label Prediction:
-Sparse Coding Tree (SCT) is an extremely fast decision tree for Millions of Labels
-Hierarchical Maximum Margin Learning for Fast Many Class prediction
-Co-hashing is hashing input data and Millions of Labels together
-Deep Discrete Prototypes for predicting Millions of Labels
-An Easy-to-hard paradigm for multi labels


Feature Selection for Big Dimensionality:
-Feature Generating Machine (FGM) for Out-of-Core Feature selection on Big Data with Trillion Dimensions
-Group Discovery Machine (GDM) for Feature Selection and Grouping on Big Data with Trillion Correlations
-Feature Selection with Performance Optimization (FSperf) for imbalanced classfication and information retrieval
-Feature Disentangling Machine (FDM) for sparse Multi-Task Feature Selection
-Sparse Perceptron Decision Tree (SPDT) is an extremely fast decision tree for Millions of Dimensions
-Doubly Approximate Nearest Neighbours on Very High Dimensional Data with Million Dimensions.
-Learning Sparse Confidence-Weighted Classifier for online feature selection on Very High Dimensional Data with Million Dimensions.
-SPARse Graph Embedding (SPARGE) framework unifies ultrahigh dimensional Feature Selection and Graph Embedding based Feature Reduction, such as sparse PCA, sparse LDA, sparse LPP, sparse SDA, etc.


Transfer Learning and Domain Adaptation:
-Transfer Component Analysis (TCA)
-Domain Adaptation Machine (DAM) is Classifier Distillation for Domain Adaptation
-Adaptive Multiple Kernel Learning (A-MKL)
-Transfer Ordinal Label Learning (TOLL) for Ordinal Regression
-Maximum Margin Target Label Learning (MMTLL)
-Sparse Heterogeneous Feature Representation (SHFR) for Multiclass Heterogeneous Transfer Learning
-Heterogeneous Feature Augmentation (HFA)
-Deep Hybrid Heterogeneous Transfer Learning (HHTL)
-Improving Generalization of Zero-shot Learning using Out-of-the-box Data
-Zero-shot Learning using Complementary Attributes
-Optimizing Performance Measures for Transfer Learning
-Transfer Learning for Clustering, Stochastic Optimization, etc
-Memetic Search with Inter-Domain Learning


Learning on Graphs:
-Graph Cross Networks with Vertex Infomax Pooling
-High-order Proximity Preserving Information Network Hashing
-Learning on Attribute-Missing Graphs
-DANA: Domain-adversarial Network Alignment
-Principal Graph and Structure Learning Based on Reversed Graph Embedding
-Long-short Distance Aggregation Networks for Positive Unlabeled Graph Learning
-Learning Image-Specific Attributes by Hyperbolic Neighborhood Graph Propagation
-Incremental Subgraph Feature Selection for Graph Classification
-Defragging Subgraph Features for Graph Classification


Crowdsourced Intelligence:
-ROPAL: Robust PL model for Rank Aggregation
-COUPLE: Rank Aggregation for any different length k-ary preferences
-Guess with Hints: A Hint-guide Mechanism for Crowdsourcing
-Nonconvex SGD optimization for classification on noisy Crowdsourced data
-Dual Filtering for Crowdsourcing Inference


Sparse Representation and Dictionary Learning:
-SC2Net: End-to-end Sparse Coding based on Differential Programming
-Laplacian Sparse Coding
-Kernel Sparse Representation
-Efficient Multi-Layer Group Sparse Coding for Concurrent Image Classification and Annotation