The Effects of Hybrid Neural Networks on Meta-Learning Objectives

Abstract

Historically, deep neural networks do not generalize well when they are trained solely on adataset/task’s objective, despite the plethora of data and computing available in the modern digital era. We propose that this is due, at least partially, to the model representations being inflexible. In this paper, we experiment with a hybrid neural network architecture that has an unsupervised model at its head (the Knowledge Representation module) and a supervised model at its tail (the Task Inference module) with the idea that we can supplement the learning of a set of related tasks with a reusable knowledge base. We analyze the two-part model in the contexts of transfer learning, few-shot learning, and curriculum learning, and train on the MNIST and SVHN datasets. The results of the experiment demonstrate that our architecture on average achieves a similar test accuracy as the end-to-end baselines, and sometimes marginally better in certain experiments depending on the sub-network combination.