Life Long Learning
catastrophic forgetting
multi-task
- computation problem ---- using all data
- storage issue ---- always keep the data
- multi-task traning can be considered as the upper bound of LLL
multi-model
- storage issue ---- can not store all models
- Knowledge cannot transfer across different models
Transfer vs LLL
- learn task2 based on the task1
- learn task2 not forgetting task1
Evaluation
Selective Synaptic Plasticity
- Basic idea: only change the unimportant parameters
- method: update the loss function
L ′ ( θ ) = L ( θ ) + λ ∑ i b i ( θ i − θ i b ) 2 L'(\theta)=L(\theta)+\lambda \sum_{i}b_i(\theta_{i}-\theta_i^b)^2 L′(θ)=L(θ)+λi∑bi(θi−θib)2
θ i b \theta_i^b θib is the model learned from the previous tasks - Gradient Episodic Memory(GEM)
Additional Neural Resource Allocation
-
Progressive Neural Networks
Add new parameters based on previous model -
PackNet
From the begining, we set a large model with enough parameters and update different parameters every task.
Memory Reply
- generating data:
We need a additional generator to generate data for the task before
- Adding new classes