Continual Backprop: Stochastic Gradient Descent with Persistent Randomness
This addresses a foundational issue in machine learning by improving the plasticity of neural networks for continual learning, which is crucial for applications requiring long-term adaptation, though it appears incremental as an extension of Backprop.
The paper tackles the problem of Backprop's performance degradation in continual learning setups, showing that initial randomness enables only initial learning but not continual adaptation. It proposes Continual Backprop, which injects random features alongside gradient descent, and demonstrates its ability to continually adapt in supervised and reinforcement learning problems with the same computational complexity as Backprop.
The Backprop algorithm for learning in neural networks utilizes two mechanisms: first, stochastic gradient descent and second, initialization with small random weights, where the latter is essential to the effectiveness of the former. We show that in continual learning setups, Backprop performs well initially, but over time its performance degrades. Stochastic gradient descent alone is insufficient to learn continually; the initial randomness enables only initial learning but not continual learning. To the best of our knowledge, ours is the first result showing this degradation in Backprop's ability to learn. To address this degradation in Backprop's plasticity, we propose an algorithm that continually injects random features alongside gradient descent using a new generate-and-test process. We call this the \textit{Continual Backprop} algorithm. We show that, unlike Backprop, Continual Backprop is able to continually adapt in both supervised and reinforcement learning (RL) problems. Continual Backprop has the same computational complexity as Backprop and can be seen as a natural extension of Backprop for continual learning.