On the Use of Deep Boltzmann Machines for Road Signs Classification

The Deep Boltzmann Machine (DBM) has been proved to be one of the most effective machine learning generative models in discriminative tasks. They've been able to overcome other generative, and even discriminative models, on relatively simple tasks, such as digits recognition, and also on more complex tasks such as simple objects recognition. However, there're only a few published results of DBM performances on other complex datasets. In this work we decided to test the efficiency of DBM, and its variant Multi-Prediction Deep Boltzmann Machine (MP-DBM), in classifying a complex dataset composed of road signs and we'll show how we've been able to train both models to reach what, at the best of our knowledge, are the best discriminative results of generative models on the road signs dataset.