Table 8 Classification accuracy of baseline algorithms with pre-trained knowledge bases limited to variously sized subsets of the training data from CORe50, Mini ImageNet and CUB

From: Building adaptive knowledge bases for evolving continual learning models

 

20% Pre-Trained

40% Pre-Trained

CLIP

CORe50

M-IN

CUB

CORe50

M-IN

CUB

MoE-Adapters7

19.64%

28.01%

25.91%

22.25%

37.43%

38.38%

DyTox6

19.25%

25.36%

27.36%

21.47%

33.96%

35.71%

Ours

20.14%

29.47%

28.41%

23.26%

39.74%

39.77%

Gain

↑2.55%

↑4.82%

↑3.84%

↑4.54%

↑6.17%

↑3.62%

 

20% Pre-Trained

40% Pre-Trained

ResNet-18

CORe50

M-IN

CUB

CORe50

M-IN

CUB

MoE-Adapters7

18.61%

26.14%

24.61%

28.74%

35.40%

37.41%

DyTox6

19.36%

23.26%

22.60%

26.25%

31.35%

33.39%

Ours

20.01%

27.52%

26.01%

28.75%

38.13%

38.05%

Gain

↑3.36

↑5.28%

↑5.69%

↑0.03%

↑7.71%

↑1.71%

  1. Classification accuracy is acquired through testing of classes that are not trained by the knowledge base.
Back to article page