Continual Semi-Supervised Learning: First International Workshop, CSSL 2021, Virtual Event, August 19–20, 2021, Revised Selected Papers (Lecture Notes in Artificial Intelligence)

Preface
Organization
Invited Talks
	Continual Learning: The Challenge
	Continual and On-the-Job Learning
	Continual Learning: Some Reflections
	Contents
International Workshop on Continual Semi-Supervised Learning: Introduction, Benchmarks and Baselines
	1 Introduction
		1.1 Contributions
		1.2 Paper Outline
	2 Related Work
		2.1 Continual Supervised Learning
		2.2 Continual Learning as Constrained Optimisation
		2.3 Unsupervised Domain Adaptation
	3 Continual Semi-Supervised Learning
		3.1 Learning
		3.2 Testing
	4 Benchmark Datasets
		4.1 Continual Activity Recognition (CAR) Dataset
		4.2 The Continual Crowd Counting (CCC) Dataset
	5 Challenges
		5.1 Protocols
		5.2 Tasks
		5.3 Baselines
	6 Results
		6.1 Results on Activity Recognition
		6.2 Results on Crowd Counting
	7 Conclusions
	References
Unsupervised Continual Learning via Pseudo Labels
	1 Introduction
	2 Related Work
		2.1 Continual Learning
		2.2 Unsupervised Representation Learning
	3 Problem Setup
	4 Our Method
		4.1 Clustering: Obtain Pseudo Label
		4.2 Incorporating into Supervised Approaches
	5 Experimental Results
		5.1 Benchmark Experimental Protocol
		5.2 Evaluation Metrics
		5.3 Implementation Detail
		5.4 Incorporating with Supervised Approaches
		5.5 Ablation Study
	6 Conclusion
	References
Transfer and Continual Supervised Learning for Robotic Grasping Through Grasping Features
	1 Introduction
	2 Problem Formulation
	3 Network Architecture
		3.1 Dataset Generation
		3.2 Supervised Transfer and Continual Learning
	4 Implementation
		4.1 Grasping Features
		4.2 Shape Decomposition
	5 Results
		5.1 Transfer and Continual Learning Pipeline
		5.2 Simulated Tests
	6 Conclusions
	References
Unsupervised Continual Learning via Self-adaptive Deep Clustering Approach
	1 Introduction
	2 Related Works
	3 Problem Formulation
	4 Learning Policy of KIERA
		4.1 Network Structure
		4.2 Structural Learning Mechanism
		4.3 Parameter Learning Mechanism
		4.4 Centroid-Based Experience Replay
	5 Proof of Concepts
		5.1 Datasets
		5.2 Implementation Notes of KIERA
		5.3 Network Structure
		5.4 Hyper-parameters
		5.5 Baseline Algorithms
		5.6 Numerical Results
	6 Conclusion
	References
Evaluating Continual Learning Algorithms by Generating 3D Virtual Environments
	1 Introduction
	2 Parametric Generation of Environments
	3 Continual Learning 3D Virtual Benchmark
	4 Examples
	5 Conclusions
	References
A Benchmark and Empirical Analysis for Replay Strategies in Continual Learning
	1 Introduction
	2 Memory Replay
		2.1 Experience Replay
		2.2 Generative Replay
	3 Evaluation
		3.1 Feature Embedding
		3.2 Model Prediction
		3.3 Maximal Interfered Retrieval (MIR)
		3.4 Datasets
	4 Results and Discussion
		4.1 Evaluation Criteria
		4.2 Replay Strategies Comparison
		4.3 Replay Data Difficulty Comparison
		4.4 Experience and Generative Replay Comparison
		4.5 Training Sequence Comparison
		4.6 Analysis
	5 Conclusion
	References
SPeCiaL: Self-supervised Pretraining for Continual Learning
	1 Introduction
	2 Related Work
	3 Background: Meta Continual Learning
	4 Proposed Method
		4.1 Generating Pseudo Labels
		4.2 Maximizing Parallel Computations
		4.3 Delayed Feature Recall
	5 Experiments
		5.1 Results
		5.2 Sparse Representation is Architecture-Dependent
	6 Conclusion
	References
Distilled Replay: Overcoming Forgetting Through Synthetic Samples
	1 Introduction
	2 Continual Learning Scenario
	3 Related Works
		3.1 Dual Memories Strategies
	4 Distilled Replay
		4.1 Buffer Distillation
		4.2 Distilled Replay Training
	5 Experiments
		5.1 Results
		5.2 Ablation Study
		5.3 Computational Times
	6 Discussion
	7 Conclusion and Future Work
	References
Self-supervised Novelty Detection for Continual Learning: A Gradient-Based Approach Boosted by Binary Classification
	1 Introduction
	2 Previous Works on Novelty Detection
	3 Proposed Methodology
		3.1 Gradient-Based Novelty Detector
		3.2 Self-supervised Binary Classifier
		3.3 Mutual Assistance Between the Binary Classifier and the Mahalanobis Path
	4 Experiments and Results
		4.1 Experimental Setup
		4.2 Batch Training of the Binary Classifier
		4.3 Performance Evaluation of Novelty Detection
		4.4 One-Class Learning with the OOD Detector
	5 Conclusion
	References
Author Index