Activation Fingerprint
A compact representation constructed from statistics of a model's intermediate-layer activations on representative inputs, such as means and variances. It characterizes behavioral changes of a model relative to its base model and can serve as the conditioning input to a hypernetwork. Because it summarizes activation statistics rather than full parameters, it provides a low-dimensional description of how a model's internal behavior has shifted.
Timeline (last 90 days)
HyperRepair learns a hypernetwork that uses activation fingerprints to summarize the model's changes relative to the pretrained base and generates a customized Repair Side Network (RSN).
An activation fingerprint refers to a compact representation formed from statistics of a model's intermediate-layer activations on a number of representative inputs, used to characterize what changes a model has undergone relative to its base model.
All Facts (2)
HyperRepair learns a hypernetwork that uses activation fingerprints to summarize the model's changes relative to the pretrained base and generates a customized Repair Side Network (RSN).
50%UnverifiedAn activation fingerprint refers to a compact representation formed from statistics of a model's intermediate-layer activations on a number of representative inputs, used to characterize what changes a model has undergone relative to its base model.
50%