Memory Aware Synapses: Learning what (not) to forget

TL;DR

The paper tackles catastrophic forgetting in lifelong learning under fixed model capacity by introducing Memory Aware Synapses (MAS), an unsupervised, online method that estimates per-parameter importance from output sensitivity. A regularizer then protects important parameters when learning new tasks, enabling selective forgetting and adaptation to unlabeled data, including test conditions. A local variant ties MAS to Hebbian learning, and MAS can be applied post-hoc to pretrained networks. Empirical results on sequential object recognition and fact embedding demonstrate state-of-the-art forgetting mitigation, with strong adaptability to data distributions without requiring labeled past data.

Abstract

Humans can learn in a continuous manner. Old rarely utilized knowledge can be overwritten by new incoming information while important, frequently used knowledge is prevented from being erased. In artificial learning systems, lifelong learning so far has focused mainly on accumulating knowledge over tasks and overcoming catastrophic forgetting. In this paper, we argue that, given the limited model capacity and the unlimited new information to be learned, knowledge has to be preserved or erased selectively. Inspired by neuroplasticity, we propose a novel approach for lifelong learning, coined Memory Aware Synapses (MAS). It computes the importance of the parameters of a neural network in an unsupervised and online manner. Given a new sample which is fed to the network, MAS accumulates an importance measure for each parameter of the network, based on how sensitive the predicted output function is to a change in this parameter. When learning a new task, changes to important parameters can then be penalized, effectively preventing important knowledge related to previous tasks from being overwritten. Further, we show an interesting connection between a local version of our method and Hebb's rule,which is a model for the learning process in the brain. We test our method on a sequence of object recognition tasks and on the challenging problem of learning an embedding for predicting $<$subject, predicate, object$>$ triplets. We show state-of-the-art performance and, for the first time, the ability to adapt the importance of the parameters based on unlabeled data towards what the network needs (not) to forget, which may vary depending on test conditions.

Figures (18)

• Figure 1: Our continuous learning setup. As common in the LLL literature, tasks are learned in sequence, one after the other. If, in between learning tasks, the agent is active and performs the learned tasks, we can use these unlabeled samples to update importance weights for the model parameters. Data that appears frequently, will have a bigger contribution. This way, the agent learns what is important and should not be forgotten.
• Figure 2: zenke2017improvedkirkpatrick2016overcoming estimate the parameters importance based on the loss, comparing the network output (light blue) with the ground truth labels (green) using training data (in yellow) (a). In contrast, we estimate the parameters importance, after convergence, based on the sensitivity of the learned function to their changes (b). This allows using additional unlabeled data points (in orange). When learning a new task, changes to important parameters are penalized, the function is preserved over the domain densely sampled in (b), while adjusting not important parameters to ensure good performance on the new task (c).
• Figure 3: Gradients flow for computing the importance weight. Local considers the gradients of each layer independently.
• Figure 4: Overall memory requirement for each method at each step of the sequence.
• Figure 5: \ref{['fig:8task_acc']} performance on each task, in accuracy, at the end of 8 tasks object recognition sequence. \ref{['fig:8task_drop']} drop in each task relative to the performance achieved after training each task.