Guided Random Projection: A Lightweight Feature Representation for Image Classification

Modern neural networks [e.g., Deep Neural Networks (DNNs)] have recently gained increasing attention for visible image classification tasks. Their success mainly results from capabilities in learning a complex feature mapping of inputs (i.e., feature representation) that carries images manifold structure relevant to the task. Despite the current popularity of these techniques, they are training-costly with Back-propagation (BP) based iteration rules. Here, we advocate a lightweight feature representation framework termed as Guided Random Projection (GRP), which is closely related to the classical random neural networks and randomization-based kernel machines. Specifically, we present an efficient optimization method that explicitly learns the distribution of random hidden weights instead of time-consuming fine-tuning or task-independent randomization configurations. Further, we also report the detailed mechanisms of the GRP with subspace theories. Experiments were conducted on visible image classification benchmarks to evaluate our claims. It shows that the proposed method achieves reasonable accuracy improvement (more than 2%) with moderate training cost (seconds level) compared with other randomization methods.