Life Long Learning(李宏毅课程)

catastrophic forgetting

multi-task

1. computation problem ---- using all data
2. storage issue ---- always keep the data
3. multi-task traning can be considered as the upper bound of LLL

multi-model

1. storage issue ---- can not store all models
2. Knowledge cannot transfer across different models

Transfer vs LLL

1. learn task2 based on the task1
2. learn task2 not forgetting task1

Evaluation

1. 

Selective Synaptic Plasticity

1. Basic idea: only change the unimportant parameters
2. method: update the loss function
   $$L^{\prime }(\theta )=L(\theta )+\lambda \sum _i{b}_i({\theta }_i-{\theta }_i^b{)}^2$$
   ${\theta }_i^b$ is the model learned from the previous tasks
3. Gradient Episodic Memory(GEM)
   

Additional Neural Resource Allocation

1. Progressive Neural Networks
   
   Add new parameters based on previous model

2. PackNet
   From the begining, we set a large model with enough parameters and update different parameters every task.
   

Memory Reply

1. generating data:
   We need a additional generator to generate data for the task before
   
2. Adding new classes