Local Sensitive Hashing (LSH) and Convolutional Neural Networks (CNNs) for Object Recognition

Mehdi Ghayoumi, Miguel Gomez, Kate E. Baumstein, Narindra Persaud, Andrew J. Perlowin

Abstract: Having a large dataset of labeled samples is necessary for the supervised training of most convolutional neural network (CNN) models. Lacking sufficient data or labeled samples for training a CNN can be problematic. To address this issue, we present a new approach for unsupervised learning that all CNN models with an image data type will be able to deploy. We tested this approach for object recognition on two popular datasets (CIFAR-10, and STL-10), and compared the results with results from available methods [11,20]. The experimental results demonstrate that our approach is comparable with the other methods of unsupervised training.