In order to contrast the explosion in size of state-of-the-art machine
learning models that can be attributed to the empirical advantages of
over-parametrization, and due to the necessity of deploying fast, sustainable,
and private on-device models on resource-constrained devices, the