Abstract: Object detection and image classification involve different kinds of features like HoG, BoW, etc. From the aspect of multi-modal learning, these tasks can be viewed as learning with different channels of features. However, in concrete applications, different modal always use different features and the extraction process of each modal’s feature costs lots of time. This makes most learning models cannot be applied in particular situations (e.g. on mobile devices, search engine which faces large scale data, etc.). It usually the case that strong modal which has a good accuracy tends to use costly features, and weak modal which works fast, yet could with worse performance. This article introduces a new multi-modal learning method, which incorporates the informative strong modal to help the weak modal learning by minimizing the prediction gap of unlabeled data between the two models. In the training phase, both strong and weak modal are trained, and the weak modal is adjusted to have a similar prediction as the strong modal on a large amount of unlabeled data. In the test phase, only the weak modal’s feature (i.e. feature with low extraction cost) is needed. Our experiments on INRIA person and caltech101 show that the proposed method works efficiently and effectively on common computer vision tasks, and with the plenty of unlabeled data, weak modal can even outperform the strong modal in some cases