Generalisation in Neural Networks Does not Require Feature Overlap
This work addresses a fundamental limitation in neural network theory for researchers and practitioners, showing that symmetry-based architectures can enable generalization without feature overlap, though it is incremental in extending convolutional principles to non-image domains.
The paper challenges the assumption that neural networks require overlapping features between training and test data to generalize, demonstrating that convolutional architectures can succeed on tasks where test features are absent from training data, such as learning identity functions and word sequence rules.
That shared features between train and test data are required for generalisation in artificial neural networks has been a common assumption of both proponents and critics of these models. Here, we show that convolutional architectures avoid this limitation by applying them to two well known challenges, based on learning the identity function and learning rules governing sequences of words. In each case, successful performance on the test set requires generalising to features that were not present in the training data, which is typically not feasible for standard connectionist models. However, our experiments demonstrate that neural networks can succeed on such problems when they incorporate the weight sharing employed by convolutional architectures. In the image processing domain, such architectures are intended to reflect the symmetry under spatial translations of the natural world that such images depict. We discuss the role of symmetry in the two tasks and its connection to generalisation.