DER: Dynamically Expandable Representation for Class Incremental Learning
We address the problem of class incremental learning, which is a core step towards achieving adaptive vision intelligence. In particular, we consider the task setting of incremental learning with limited memory and aim to achieve better stability-plasticity trade-off. To this end, we propose a novel two-stage learning approach that utilizes a dynamically expandable representation for more effective incremental concept modeling. Specifically, at each incremental step, we freeze the previously learned representation and augment it with additional feature dimensions from a new learnable feature extractor. This enables us to integrate new visual concepts with retaining learned knowledge. We dynamically expand the representation according to the complexity of novel concepts by introducing a channel-level mask-based pruning strategy. Moreover, we introduce an auxiliary loss to encourage the model to learn diverse and discriminate features for novel concepts. We conduct extensive experiments on the three class incremental learning benchmarks and our method consistently outperforms other methods with a large margin.
PDF Abstract CVPR 2021 PDF CVPR 2021 AbstractCode
Results from the Paper
Task | Dataset | Model | Metric Name | Metric Value | Global Rank | Benchmark |
---|---|---|---|---|---|---|
Incremental Learning | CIFAR-100 - 50 classes + 10 steps of 5 classes | DER(Modified ResNet-32) | Average Incremental Accuracy | 66.36 | # 7 | |
Incremental Learning | CIFAR-100 - 50 classes + 10 steps of 5 classes | DER(Standard ResNet-18) | Average Incremental Accuracy | 72.45 | # 3 | |
Incremental Learning | CIFAR-100 - 50 classes + 2 steps of 25 classes | DER (w/o P) | Average Incremental Accuracy | 74.61 | # 3 | |
Incremental Learning | CIFAR-100 - 50 classes + 5 steps of 10 classes | DER(Standard ResNet-18) | Average Incremental Accuracy | 72.60 | # 3 | |
Incremental Learning | CIFAR-100 - 50 classes + 5 steps of 10 classes | DER(Modified Res-32) | Average Incremental Accuracy | 67.60 | # 7 | |
Incremental Learning | CIFAR100-B0(10steps of 10 classes) | DER(ResNet-18) | Average Incremental Accuracy | 74.64 | # 3 | |
Incremental Learning | CIFAR100B020Step(5ClassesPerStep) | DER(ResNet-18) | Average Incremental Accuracy | 73.98 | # 3 | |
Incremental Learning | CIFAR100B050S(2ClassesPerStep) | DER(ResNet-18) | Average Incremental Accuracy | 72.05 | # 1 | |
Incremental Learning | CIFAR-100-B0(5steps of 20 classes) | DER(w/o P) | Average Incremental Accuracy | 76.80 | # 3 | |
Incremental Learning | ImageNet100 - 10 steps | DER | Average Incremental Accuracy | 76.12 | # 8 | |
Final Accuracy | 66.07 | # 5 | ||||
Average Incremental Accuracy Top-5 | 92.79 | # 4 | ||||
Final Accuracy Top-5 | 88.38 | # 2 | ||||
Incremental Learning | ImageNet100 - 10 steps | DER w/o Pruning | Average Incremental Accuracy | 77.18 | # 6 | |
Final Accuracy | 66.70 | # 4 | ||||
Average Incremental Accuracy Top-5 | 93.23 | # 3 | ||||
Final Accuracy Top-5 | 87.52 | # 5 | ||||
# M Params | 112.27 | # 7 | ||||
Incremental Learning | ImageNet-100 - 50 classes + 10 steps of 5 classes | DER | Average Incremental Accuracy | 77.73 | # 2 | |
Incremental Learning | ImageNet - 10 steps | DER w/o Pruning | Average Incremental Accuracy | 68.84 | # 3 | |
Final Accuracy | 60.16 | # 2 | ||||
Average Incremental Accuracy Top-5 | 88.17 | # 2 | ||||
Final Accuracy Top-5 | 82.86 | # 2 | ||||
# M Params | 116.89 | # 6 | ||||
Incremental Learning | ImageNet - 10 steps | DER | Average Incremental Accuracy | 66.73 | # 6 | |
Final Accuracy | 58.62 | # 3 | ||||
Average Incremental Accuracy Top-5 | 87.08 | # 3 | ||||
Final Accuracy Top-5 | 81.89 | # 3 |