On the benecial effects of reinjections for continual learning
Abstract
Deep learning consistently delivers remarkable results in a wide range of
applications, but artificial neural networks still suffer from catastrophic
forgetting of old knowledge as new knowledge is learned. Rehearsal
methods overcome catastrophic forgetting by replaying an amount of pre-
viously learned data stored in dedicated memory buffers. On the other
hand, pseudo-rehearsal methods generate pseudo-samples to emulate pre-
viously learned data, alleviating the need for dedicated buffers. This
paper █first shows how it is possible to alleviate catastrophic forgetting
with a pseudo-rehearsal method without employing memory buffers or
generative models to generate the pseudo-samples. We propose a hybrid
architecture similar to that of an autoencoder with additional neurons
to classify the input. This architecture preserves specific properties of
autoencoders by allowing the generation of pseudo-samples through a
sampling procedure with random noise and reinjection (i.e. iterative
sampling). The generated pseudo-samples are then interwoven with the
new examples to acquire new knowledge without forgetting the previous
ones. Secondly, we combine the two methods (rehearsal and pseudo-
rehearsal) in the hybrid architecture. Examples stored in small memory
buffers are used as seeds instead of noise to improve the process of
generating pseudo-samples and retrieving previously learned knowledge.
We demonstrate that reinjections are suitable for rehearsal and pseudo-
rehearsal approaches and show state-of-the-art results on rehearsal
methods for small buffer sizes. We evaluate our method extensively
on MNIST, CIFAR-10 and CIFAR-100 image classi█cation datasets,
and present state-of-the-art performance using tiny memory buffers.
Origin : Files produced by the author(s)