Continuous learning, lifelong learning, episodic memory, memory module paper summary -- gradient episodic memory promotes continuous learning

Gradient Episodic Memory for Continual Learning

Gradient episodic memory promotes continuous learning

1. Related information ：

• Company ：Facebook AI research
• Time ：2017
• Code ：github Mirror address
• The paper ：csdn Download address

1. Paper objectives ：

For a large number of tasks 、 Continuous learning of a small number of samples , stay MNIST and CIFAR-100 Experiments on data sets

2. Important details ：

• Presupposition ： Within each task , Data is independent and distributed
• The data shows ： A triple ： Input , Task identifier 、 Output
• Forward migration ： Learn a new task , Impact on future learning follow-up tasks
• Backward migration ： Learn a new task , Impact on previously learned tasks
• Episodic memory module ： Store data pairs for past tasks
• Continuous learning realizes ： All tasks share parameters , Ensure that when learning new tasks and updating parameters , The loss of past task data in memory does not increase , So as to solve catastrophic forgetting
• Measure loss changes ： Measured by the included angle of gradient direction , Cosine value is greater than or equal to 0 In the same direction , It will not affect the performance of past tasks , Have good backward migration ability
• Gradient update direction correction ： The selection is closest to the original parameter update （ use L2 Norm measures distance ）, And meet the projection of measuring loss change conditions
• Quadratic programming and its dual problem

3. Inspire and think ：

• Add task descriptor ： More detailed task description . Measure the similarity between different tasks through task descriptors , Infer the relationship between tasks , Clear the ambiguity between similar tasks
  • High similarity ： Learning new tasks can speed up learning , Even realize 0 Second study
  • Low similarity ： Focus on backward migration , Measure loss changes , It is not necessary to measure the loss change of tasks with high similarity
• Advanced memory management ： Build a core set of past tasks