Dealing with Unknowns in Continual Learning for End-to-end Automatic Speech Recognition

Learning continually from data is a
task executed effortlessly by humans but remains to be of significant
challenge for machines. Moreover, when encountering unknown test
scenarios machines fail to generalize. We propose a mathematically
motivated dynamically expanding end-to-end model of independent
sequence-to-sequence components trained on different data sets that
avoid catastrophically forgetting knowledge acquired from previously
seen data while seamlessly integrating knowledge from new data. During
inference, the likelihoods of the unknown test scenario are computed
using internal model activation distributions. The inference made by
each independent component is weighted by the normalized likelihood
values to obtain the final decision.