Visualization of high dimensional image features for classification

In image classification tasks, the image is rarely represented as only a collection of raw pixels. Myriad alternative representations, from Gaussian kernels to bags-of-words to layers of a convolutional neural network, have been proposed both to decrease the dimensionality of the task and, more importantly, to move into a space which better facilitates classification. This work explores several methods for evaluating the suitability of the high-dimensional image representations produced by the layers of a convolutional neural network. Quantitative methods are considered along with visualization techniques. These visualizations and metrics aid in understanding what classifiers might be better suited to a given dataset. We demonstrate these methods on the MNIST handwritten digit dataset and compare the results with accuracies obtained from classifiers.