Direct classification from compressively sensed images via deep Boltzmann machine

We examine a potential technique of performing a classification task based on compressively sensed (CS) data, skipping a computationally expensive reconstruction step. A deep Boltzmann machine is trained on a compressive representation of MNIST handwritten digit data, using a random orthoprojector sensing matrix. The network is first pre-trained on uncompressed data in order to learn the structure of the dataset. The outer network layers are then optimized using backpropagation. We find this approach achieves a 1.21% test data error rate at a sensing rate of 0.4, compared to a 0.99% error rate for non-compressive data.