We apply deep belief networks of restricted Boltzmann machines to bags of words of sift features obtained from databases of 13 Scenes, 15 Scenes and Caltech 256 and study experimentally their behavior and performance. We find that the final performance in the supervised phase is reached much faster if the system is pre-trained. Pre-training the system on a larger dataset keeping the supervised dataset fixed improves the performance (for the 13 Scenes case). After the unsupervised pre-training, neurons arise that form approximate explicit representations for several categories (meaning they are mostly active for this category). The last three facts suggest that unsupervised training really discovers structure in these data. Pre-training can be done on a completely different dataset (we use Corel dataset) and we find that the supervised phase performs just as good (on the 15 Scenes dataset). This leads us to conjecture that one can pre-train the system once (e.g. in a factory) and subsequently apply it to many supervised problems which then learn much faster. The best performance is obtained with single hidden layer system suggesting that the histogram of sift features doesn't have much high level structure. The overall performance is almost equal, but slightly worse then that of the support vector machine and the spatial pyramidal matching.