On the benecial effects of reinjections for continual learning

Deep learning consistently delivers remarkable results in a wide range ofapplications, but artificial neural networks still suffer from catastrophicforgetting of old knowledge as new knowledge is learned. Rehearsalmethods overcome catastrophic forgetting by replaying an amount of pre-viously learned data stored in dedicated memory buffers. On the otherhand, pseudo-rehearsal methods generate pseudo-samples to emulate pre-viously learned data, alleviating the need for dedicated buffers. Thispaper █first shows how it is possible to alleviate catastrophic forgettingwith a pseudo-rehearsal method without employing memory buffers orgenerative models to generate the pseudo-samples. We propose a hybridarchitecture similar to that of an autoencoder with additional neuronsto classify the input. This architecture preserves specific properties ofautoencoders by allowing the generation of pseudo-samples through asampling procedure with random noise and reinjection (i.e. iterativesampling). The generated pseudo-samples are then interwoven with thenew examples to acquire new knowledge without forgetting the previousones. Secondly, we combine the two methods (rehearsal and pseudo-rehearsal) in the hybrid architecture. Examples stored in small memorybuffers are used as seeds instead of noise to improve the process ofgenerating 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 rehearsalmethods for small buffer sizes. We evaluate our method extensivelyon MNIST, CIFAR-10 and CIFAR-100 image classi█cation datasets,and present state-of-the-art performance using tiny memory buffers.