I think a working continual learning implementation would mess with convergence-based results which have a relatively fixed training data distribution and only modify the starting seeds. This is mostly because a continual learning system is constantly drifting “off the base distribution” and incorporating new data. In other words, the car model has seen data from places and distributions the attacker’s base model never will.
This seems right to me. But maybe the drift in distribution mainly affects certain parameters, and the divergence tokens affect a separate set of parameters (in early layers) s.t. the downstream effect still persists even after being in OOD
From my understanding from this paper, lottery tickets are invariant to optimiser, datatype, and other model properties (in this experimental setting), suggesting lottery tickets encode some basic properties of the task.
It seems unlikely lottery tickets based on fundamental task properties would change with continual learning without other problems emerging (catastrophic forgetting).
I think a working continual learning implementation would mess with convergence-based results which have a relatively fixed training data distribution and only modify the starting seeds. This is mostly because a continual learning system is constantly drifting “off the base distribution” and incorporating new data. In other words, the car model has seen data from places and distributions the attacker’s base model never will.
This seems right to me. But maybe the drift in distribution mainly affects certain parameters, and the divergence tokens affect a separate set of parameters (in early layers) s.t. the downstream effect still persists even after being in OOD
From my understanding from this paper, lottery tickets are invariant to optimiser, datatype, and other model properties (in this experimental setting), suggesting lottery tickets encode some basic properties of the task.
It seems unlikely lottery tickets based on fundamental task properties would change with continual learning without other problems emerging (catastrophic forgetting).