Continual learning with invertible generative models.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Catastrophic forgetting (CF) is a major challenge in continual learning, where neural networks overwrite previously acquired knowledge when trained on new tasks.
• Existing methods to mitigate CF include weight regularization and rehearsal strategies, with generative models used to create synthetic data for rehearsal.

Purpose of the Study:

• To propose a novel method that integrates regularization and generative-based rehearsal to combat catastrophic forgetting in neural networks.
• To develop a memory-efficient approach with constant memory overhead throughout the training process.

Main Methods:

• A generative model based on a normalizing flow (NF), a probabilistic and invertible neural network, is trained on the internal embeddings of the neural network.
• The invertibility of the NF is leveraged to implement a straightforward regularization technique for network embeddings concerning past tasks.
• The proposed method combines the strengths of regularization and generative-based rehearsal strategies.

Main Results:

• The proposed method demonstrates favorable performance compared to state-of-the-art approaches in the literature.
• The memory overhead is maintained at a constant level by utilizing a single NF throughout training.
• The approach achieves this with bounded computational power and memory requirements.

Conclusions:

• The novel method effectively addresses catastrophic forgetting by synergistically combining regularization and generative rehearsal.
• The use of a normalizing flow offers an efficient and scalable solution for continual learning.
• This technique provides a promising direction for developing more robust and persistent neural network models.