A Experiment Details on Representation
–Neural Information Processing Systems
We follow the same optimization step of SimSiam [11]. ResNet-18 [22] with the last fc layer removed. We use the same projection and prediction networks in SimSiam. 's input and output is After pretraining, the representation encoder f is finetuned for downstream tasks. We use the symmetrized cosine similarity loss from SimSiam.
Neural Information Processing Systems
Nov-17-2025, 05:50:20 GMT