The technology of single-pixel imaging extends visualization capabilities beyond pixel-matrix-based devices. One of possible applications for this technology is fast classification of objects without the need for reconstruction of an  image. The single-pixel camera gathers light statistics and then a computational algorithm – such as a neural network – decides on what is the object been illuminated. We train a convolutional neural network on simulated data from single-pixel camera and demonstrate effectiveness of classification images of handwritten digits.