An Exemplar-Informed Approach to Speech and Language Tasks

The field of machine learning has drawn heavily from the fields of psychology and neuroscience, in particular in the development of artificial neural network architectures, which are based on simplified versions of structures in the brain. While effective for many tasks, neural networks do not, in general, incorporate any way of storing specific experiences, instead using training data to parameterise a model, and then discarding the training date prior to inference. We explore an alternative option: a simple, explainable model from the field of human psychology called Minerva 2, which uses previously seen examples to perform classification or regression. By comparing Minerva 2 with neural networks, we demonstrate that Minerva 2 is in fact a neural network itself, with parameters taken directly from the data, rather than being trained by backpropagation. We propose new architectures, which are based on Minerva 2 and incorporate both a memory of previous examples and parameterisation that allows model flexibility. We show that feature representation is crucial for this type of model, which might explain the lack of representation of this type of model in the literature. Speech and text representations have improved rapidly in recent years, however, and if this trend continues, simple, interpretable models such those proposed here will become more competitive. As evidence of this, we use high quality speech representations in conjunction with a Minerva-based model to demonstrate state-of-the-art performance on a speech intelligibility task.