Meta-Learning Representation for Continual Learning

NeurIPS 2019 8-24-2020

Motivation

Neural Networks trained online on a correlated stream of data suffer from catastrophic forgeting. We propose learning a representation that is robust to forgeting. To learn the representation, we propose OML, a second-order meta-learning objective that directly minimizes interference. Highly sparse representations naturally emerge by minimizing our proposed objective.

Two parts in the model as shown from the figure:

$\phi_{\theta}$ and $g_w$

two parts: RLN and PLN

Meta-parameters: A deep neural network that transforms high-dimensional input data to a representation $\mathcal{R}^d$ which is more conducive for continual learning

Adaptation parameters: A simple neural network that learns continually from $\mathcal{R}^d$

Meta-Training:

Step 1: adaptation (Inner loop updates only for $g_w$ )

Step 2: compute Meta-Loss on the complete task dataset (similar to MAML)

Step 3: Meta-update: differentiating meta-loss through the adaptation phase (for $\phi_\theta$ and $g_w$ ) [similar to MAML]

Meta-Testing:

Step 1: Adaptation (Inner loop updates only for $g_w$ , use $\phi_\theta$ and $g_w$ learnt from Meta-training)

Step 2: Evaluation (fix $\phi_\theta$ and $g_w$, to compute the accuracy)

Note:

• check online meta learning paper again, compare the difference

• attention: domain shift problem

• two parts architectures: it looks very common in many papers

Reference:

• https://papers.nips.cc/paper/8458-meta-learning-representations-for-continual-learning

• https://www.youtube.com/watch?v=XlEqFeQiuhk