Import recent advances in the domain of incremental or continual learning with DNNs, such as Elastic Weight Consolidation (EWC) or Incremental Moment Matching (IMM) rely on a quantity termed the Fisher information matrix (FIM). While the results obtained in this way are very promising, the use of the FIM relies on the assumptions that (a) the FIM can be approximated by its diagonal, and (b) that FIM diagonal entries are related to the variance of a DNN parameter in the context of Bayesian neural networks. In addition, the FIM is notoriously difficult to compute in automatic differentiation (AD) systems frameworks like TensorFlow, and existing implementations require an excessive use of memory due to this problem. We present the Matrix of SQuares (MaSQ), computed similarly as the FIM, but whose use in EWC-like algorithms follows directly from the calculus of derivatives and requires no additional assumptions. Additionally, MaSQ computation in AD frameworks is much simpler and more memory-efficient FIM computation. When using MaSQ together with EWC we show superior or equal performance to FIM/EWC on a variety of benchmark tasks.
展开▼
