I’m trying to perform image classification on CIFAR-10 using ResNet. By pretraining ResNet, I was able to achieve a performance improvement of 92% on CIFAR-10.

Next, I want to add a fully connected layer after ResNet to create a network that determines whether the output is true or false. I will call this network the “decision network.”

The decision network takes the final softmax output of the pretrained network as input and determines whether it is true or false.

However, the final softmax output of the pretrained network often has a maximum value of 0.99 or higher even when it misclassifies. As a result, the decision network is not able to effectively distinguish between true and false.

In such cases, how should I train the network to improve its performance?