We propose BayesCNN, a variant of CNN that uses probability distribution as weights instead of point estimates to achieve state of the art results. The paper can be found here : https://arxiv.org/abs/1806.05978
If you want to collaborate or check the code, here is the GitHub : https://github.com/kumar-shridhar/PyTorch-BayesianCNN
There are several ways in which the results can be improved. But first, we are working on improving the training time and applying the concept of BayesCNN to other complex architectures like ResNet or InceptionNet. Any comments, ideas or suggestions are welcome. Feel free to try it out and let us know your thoughts.