Imbalanced class accuracy improvement due to sleep. Each row shows experiments with data reduction for one specific class (shown on the left), with the percentage of reduction shown on the horizontal axis. Each cell shows the class-wise accuracy of the underrepresented class before sleep (top value) and after sleep (bottom value). The color map is based on the change in accuracy, ∆ = After Sleep -Before Sleep. Reds indicate a positive difference (improvement), while blues indicate a negative difference (drop in accuracy). Note, many red squares showing class-wise improvement with only a few blue squares showing class-wise performance loss.

Contexts in source publication

Context 1

... we systematically tested the effect of SRC in class imbalanced settings, i.e., in addition to training the network on limited data, one selected class was explicitly underrepresented. Figure 4 shows an example of such analysis when we used a 10% subset of MNIST training data and further limited the number of images for one selected class during training. In practice, we developed these datasets by first randomly selecting a 10% subset of the MNIST dataset and ensuring each class had the exact same number of images. ...

Context 2

... example, digit 0 showed high class-wise accuracy when more than 70-80% digit 0's were used for training (i.e., more than 7-8% of digit 0's in the total dataset), while digit 5 had very low performance even when all data were used (i.e., 10% of 5's in the total dataset). After SRC, most classes (except digit 8) showed a positive improvement in class-wise accuracy (Figure 4). The magnitude of the gain and the range of data reduction where the gain was observed were varied between digits likely because of the different sensitivity to data reduction. ...