Measuring Catastrophic Forgetting in Neural Networks

Abstract

Deep neural networks are used in many state-of-the-art systems for machine perception. Once a network is trained to do a specific task, e.g., bird classification, it cannot easily be trained to do new tasks, e.g., incrementally learning to recognize additional bird species or learning an entirely different task such as flower recognition. When new tasks are added, typical deep neural networks are prone to catastrophically forgetting previous tasks. Networks that are capable of assimilating new information incrementally, much like how humans form new memories over time, will be more efficient than re-training the model from scratch each time a new task needs to be learned. There have been multiple attempts to develop schemes that mitigate catastrophic forgetting, but these methods have not been directly compared, the tests used to evaluate them vary considerably, and these methods have only been evaluated on small-scale problems (e.g., MNIST). In this paper, we introduce new metrics and benchmarks for directly comparing five different mechanisms designed to mitigate catastrophic forgetting in neural networks: regularization, ensembling, rehearsal, dual-memory, and sparse-coding. Our experiments on real-world images and sounds show that the mechanism(s) that are critical for optimal performance vary based on the incremental training paradigm and type of data being used, but they all demonstrate that the catastrophic forgetting problem has yet to be solved.

To appear in AAAI 2018

Cited by in corpus (75)

• A continual learning survey: Defying forgetting in classification tasks
• Adaptive Aggregation Networks for Class-Incremental Learning
• Re-evaluating Continual Learning Scenarios: A Categorization and Case for Strong Baselines
• One Detector to Rule Them All: Towards a General Deepfake Attack Detection Framework
• Continual Learning in Sensor-based Human Activity Recognition: an Empirical Benchmark Analysis
• DualNet: Continual Learning, Fast and Slow
• Domain-Incremental Continual Learning for Mitigating Bias in Facial Expression and Action Unit Recognition
• Pseudo-Recursal: Solving the Catastrophic Forgetting Problem in Deep Neural Networks
• Orthogonal Gradient Descent for Continual Learning
• An Investigation of Replay-based Approaches for Continual Learning
• SpaRCe: Improved Learning of Reservoir Computing Systems through Sparse Representations
• Continual Learning with Gated Incremental Memories for sequential data processing
• Defining Benchmarks for Continual Few-Shot Learning
• Continual Learning with Adaptive Weights (CLAW)
• Improved Schemes for Episodic Memory-based Lifelong Learning
• Self-Supervised Training Enhances Online Continual Learning
• Replay in Deep Learning: Current Approaches and Missing Biological Elements
• Anatomy of Catastrophic Forgetting: Hidden Representations and Task Semantics
• Ekya: Continuous Learning of Video Analytics Models on Edge Compute Servers
• Preserving Earlier Knowledge in Continual Learning with the Help of All Previous Feature Extractors
• General Value Function Networks
• REMIND Your Neural Network to Prevent Catastrophic Forgetting
• Beneficial Perturbation Network for designing general adaptive artificial intelligence systems
• An Adaptive Random Path Selection Approach for Incremental Learning
• The Next Big Thing(s) in Unsupervised Machine Learning: Five Lessons from Infant Learning
• Exploring Benefits of Transfer Learning in Neural Machine Translation
• Quantum Continual Learning Overcoming Catastrophic Forgetting
• Continual Learning in Recurrent Neural Networks
• Orchestrating the Development Lifecycle of Machine Learning-Based IoT Applications: A Taxonomy and Survey
• Lifelong Machine Learning with Deep Streaming Linear Discriminant Analysis
• Linear Mode Connectivity in Multitask and Continual Learning
• Continual Learning in Human Activity Recognition: an Empirical Analysis of Regularization
• Tackling Catastrophic Forgetting and Background Shift in Continual Semantic Segmentation
• Do Deep Neural Networks Forget Facial Action Units? -- Exploring the Effects of Transfer Learning in Health Related Facial Expression Recognition
• Transient Non-Stationarity and Generalisation in Deep Reinforcement Learning
• Variational Prototype Replays for Continual Learning
• CoLLIE: Continual Learning of Language Grounding from Language-Image Embeddings
• Never Forget: Balancing Exploration and Exploitation via Learning Optical Flow
• Dissecting Catastrophic Forgetting in Continual Learning by Deep Visualization
• Autoencoder-Based Incremental Class Learning without Retraining on Old Data
• Memory Efficient Experience Replay for Streaming Learning
• Learning Sparse Representations Incrementally in Deep Reinforcement Learning
• Memory-Efficient Semi-Supervised Continual Learning: The World is its Own Replay Buffer
• Neural Network Retraining for Model Serving
• Answer Them All! Toward Universal Visual Question Answering Models
• A Self-Supervised Feature Map Augmentation (FMA) Loss and Combined Augmentations Finetuning to Efficiently Improve the Robustness of CNNs
• FReTAL: Generalizing Deepfake Detection using Knowledge Distillation and Representation Learning
• Continual Learning for Natural Language Generation in Task-oriented Dialog Systems
• Continual Learning with Fully Probabilistic Models
• Thalamocortical contribution to solving credit assignment in neural systems
• Learning Invariant Representation for Continual Learning
• Towards a practical measure of interference for reinforcement learning
• Fuzzy Tiling Activations: A Simple Approach to Learning Sparse Representations Online
• CLEVA-Compass: A Continual Learning EValuation Assessment Compass to Promote Research Transparency and Comparability
• Selective Forgetting of Deep Networks at a Finer Level than Samples
• The State of Lifelong Learning in Service Robots: Current Bottlenecks in Object Perception and Manipulation
• Accretionary Learning with Deep Neural Networks
• A Procedural World Generation Framework for Systematic Evaluation of Continual Learning
• Towards Recognizing New Semantic Concepts in New Visual Domains
• Explaining How Deep Neural Networks Forget by Deep Visualization
• Selective Replay Enhances Learning in Online Continual Analogical Reasoning
• Learning in Deep Neural Networks Using a Biologically Inspired Optimizer
• Class-Incremental Experience Replay for Continual Learning under Concept Drift
• Seq2Biseq: Bidirectional Output-wise Recurrent Neural Networks for Sequence Modelling
• SEER: Performance-Aware Leader Election in Single-Leader Consensus
• On Memory Mechanism in Multi-Agent Reinforcement Learning
• How saccadic vision might help with theinterpretability of deep networks
• Using Fictitious Class Representations to Boost Discriminative Zero-Shot Learners
• Recent Advances in Large Margin Learning
• Continuous Learning for Large-scale Personalized Domain Classification
• Plug-and-Play Anomaly Detection with Expectation Maximization Filtering
• Class Incremental Online Streaming Learning
• Bridge Networks: Relating Inputs through Vector-Symbolic Manipulations
• Learning to Continually Learn Rapidly from Few and Noisy Data
• Measuring Information Transfer in Neural Networks