Continual Learning
Introduction to Continual Learning / LifeLong Learning / Incremental Learning in image classification, object detection.
Datasets
CIFAR-10/CIFAR-100
CIFAR-10
- 50000 training images
- 10000 test images
- 10 classes of 32x32 color images
CIFAR-100
- 500 training images per class
- 100 test images per class
- 100 classes of 32x32 color images
CORe50
CORe50, specifically designed for Continual Object Recognition, is a collection of 50 domestic objects belonging to 10 categories: plug adapters, mobile phones, scissors, light bulbs, cans, glasses, balls, markers, cups and remote controls.
Classification can be performed at object level (50 classes) or at category level (10 classes).
SAILenv
Paper: Evaluating Continual Learning Algorithms by Generating 3D Virtual Environments
- Pixel-wise Annotations
- Object Library
-
Ready-To-Go Scenes
- Server Executables (Sample Scenes): Version Directory
- Source Unity Project (Customizable): Source Code
- Client Python API: Source Code (GitHub), Pip Package
- 3D Models .OBJ for Adversarial Attacks: ZIP archive
from sailenv.agent import Agent
agent = Agent(width=256, height=192, host="192.168.1.3", port=8085)
agent.register()
agent.change_scene(agent.scenes[2])
while True:
frame_views = agent.get_frame()
...
agent.delete()
Continual Learning
Blog: 李宏毅 lifelong learning
Blog: Catastrophic Forgetting in Neural Networks Explained
Colab: An example of catastrophic forgetting in PyTorch
The effect of AR1 (CwR+Syn) is displayed in the figure below, based on the CORe50 dataset.
LWF (Learning Without Forgetting)
Paper: Learning without Forgetting (LWF)
EWC (Elastic Weights Consolidation)
Paper: Overcoming catastrophic forgetting in neural networks
Paper: Elastic Weight Consolidation (EWC): Nuts and Bolts
Code: https://github.com/ariseff/overcoming-catastrophic
CWR (Copy Weights with Re-init), CWR+
Paper: Continuous Learning in Single-Incremental-Task Scenarios
GDM (Growing Dual Memory)
Paper: Lifelong Learning of Spatiotemporal Representations with Dual-Memory Recurrent Self-Organization
Code: https://github.com/giparisi/GDM
Toolbox: GWR Toolbox
AR1
Paper: Latent Replay for Real-Time Continual Learning
Code: AR1* with Latent Replay
CAT
Paper: Continual Learning of a Mixed Sequence of Similar and Dissimilar Tasks
Code: https://github.com/ZixuanKe/CAT
GDM for Lifelong 3D Object Recognition
Class-Incremtnatl Learning with Generative Classifier
Paper: Class-Incremental Learning with Generative Classifiers
Code: https://github.com/GMvandeVen/class-incremental-learning
Kaggle: https://kaggle.com/rkuo2000/class-incremental-learning
RMN (Revelance Mapping Network)
paper: Understanding Catastrophic Forgetting and Remembering in Continual Learning with Optimal Relevance Mapping
Code: https://gitlab.com/prakhark2/relevance-mapping-networks
PyContinual
PyContinual (An Easy and Extendible Framework for Continual Learning)
Paper:
- Achieving Forgetting Prevention and Knowledge Transfer in Continual Learning
- CLASSIC: Continual and Contrastive Learning of Aspect Sentiment Classification Tasks
- Adapting BERT for Continual Learning of a Sequence of Aspect Sentiment Classification Tasks
- Continual Learning with Knowledge Transfer for Sentiment Classification
- Continual Learning of a Mixed Sequence of Similar and Dissimilar Tasks
Features:
- Datasets: It currently supports Language Datasets (Document/Sentence/Aspect Sentiment Classification, Natural Language Inference, Topic Classification) and Image Datasets (CelebA, CIFAR10, CIFAR100, FashionMNIST, F-EMNIST, MNIST, VLCS)
- Scenarios: It currently supports Task Incremental Learning and Domain Incremental Learning
- Training Modes: It currently supports single-GPU. You can also change it to multi-node distributed training and the mixed precision training.
LwF-ECG
MAML (Model-Agnostic Meta-Learning)
Blog: MAML模型介绍及算法详解
Paper: Model-Agnostic Meta-Learning for Fast Adaption of Deep Networks
Code: https://github.com/cbfinn/maml
Blog: MAML复现全部细节和经验教训(Pytorch)
Code: https://github.com/miguealanmath/MAML-Pytorch
Lifelong Object Detection
Paper: Lifelong Object Detection
RECALL
Paper: RECALL: Replay-based Continual Learning in Semantic Segmentation
Contrast R-CNN
Paper: Contrast R-CNN for Continual Learning in Object Detection
This site was last updated December 22, 2022.