Given the prevalence of crowd sourced labor in creating Natural Language processing datasets, these aforementioned sets have become increasingly large. For instance, the SQUAD dataset currently sits at over 80,000 records. However, because the English language is rather repetitive in structure, the distribution of word frequencies in the SQUAD dataset's contexts are relatively unchanged. By measuring each sentences distance from the co-variate distance of frequencies of all sentences in the dataset, we identify 10,500 examples that create a more uniform distribution for training. While fine-tuning ELECTRA [4] on this subset of examples reaches better performance to a model trained on all 87,000 examples. Herein we introduce a methodology for systematically pruning datasets for fine tuning reaching better out of sample performance.
Recent developments using End-to-End Deep Learning models have been shown to have near or better performance than state of the art Recurrent Neural Networks (RNNs) on Automatic Speech Recognition tasks. These models tend to be lighter weight and require less training time than traditional RNN-based approaches. However, these models take frequentist approach to weight training. In theory, network weights are drawn from a latent, intractable probability distribution. We introduce BayesSpeech for end-to-end Automatic Speech Recognition. BayesSpeech is a Bayesian Transformer Network where these intractable posteriors are learned through variational inference and the local reparameterization trick without recurrence. We show how the introduction of variance in the weights leads to faster training time and near state-of-the-art performance on LibriSpeech-960.