RQ1

We choose the classical ResNet, as the subject model architecture. In addition, we train the models on the CIFAR-10 dataset until the model

performance converges. we design an experiment to observe the semantic architecture of a deep classifier by controlling three variables.

Model Size: ResNet-14  ResNet-44 ResNet-86 ResNet-110 ResNet-152

Dataset Size: 6000 14000 30000 46000 60000

Task Size: 2-classification 4-classification 6-classification 8-classification 10-classification

With a total of 125 (5*5*5) configurations, we train each configuration(model) (i.e., architecture-dataset-task combination)  10 times. There will be 1250(125*10) models. The hyperparameters' details are as follows:

batch size: 128           learning rate: 0.01        weight_decay: 0.005          epochs: 300        initializer: he_normal