Parameter-Efficient Deep Neural Networks With Bilinear Projections

Recent research on deep neural networks (DNNs) has primarily focused on improving the model accuracy. Given a proper deep learning framework, it is generally possible to increase the depth or layer width to achieve a higher level of accuracy. However, the huge number of model parameters imposes more computational and memory usage overhead and leads to the parameter redundancy. In this article, we address the parameter redundancy problem in DNNs by replacing conventional full projections with bilinear projections (BPs). For a fully connected layer with $D$ input nodes and $D$ output nodes, applying BP can reduce the model space complexity from $\mathcal {O}(D^{2})$ to $\mathcal {O}(2D)$ , achieving a deep model with a sublinear layer size. However, the structured projection has a lower freedom of degree compared with the full projection, causing the underfitting problem. Therefore, we simply scale up the mapping size by increasing the number of output channels, which can keep and even boosts the model accuracy. This makes it very parameter-efficient and handy to deploy such deep models on mobile systems with memory limitations. Experiments on four benchmark data sets show that applying the proposed BP to DNNs can achieve even higher accuracies than conventional full DNNs while significantly reducing the model size.