Please enable JavaScript.
Coggle requires JavaScript to display documents.
Continual Learning (Not Confirmed (Expand Neural Network Capacity can lead…
Continual Learning
Not Confirmed
Expand Neural Network Capacity can lead to Positive Forward Transfer?
Will Fisher value still eligible after some training?
What's variational continual learning - coreset mentioned in ORACLE
Assume steep theta likelihood equals steep loss function
Positive Forward Transfer
Expand capacity
ORACLE style
Progressive Net of P&C
External Memory - Probably only solve partially
Technique Options / or some Random Ideas
EWC to select Experience
Use Fisher value to determine weather expand NN capacity
Eliminate Catastrophic Forgetting
CLEAR
Save all experience
V-trace handles off-policy errors
Behavior cloning
Useful Experience:
Smallest set of experience that after learning can lead to maximum rewards
Assume low TD Error equals high rewards
-> not really it only indicates given current state (not all future state) it's a good Q value estimate
Need to find the experience after learning can lead to lowest value loss
Positive Backward Transfer