Deep Neural Networks are powerful models that attained remarkable results on\na variety of tasks. These models are shown to be extremely efficient when\ntraining and test data are drawn from the same distribution. However, it is not\nclear how a network will act when it is fed with an out-of-distribution\nexample. In this work, we consider the problem of out-of-distribution detection\nin neural networks. We propose to use multiple semantic dense representations\ninstead of sparse representation as the target label. Specifically, we propose\nto use several word representations obtained from different corpora or\narchitectures as target labels. We evaluated the proposed model on computer\nvision, and speech commands detection tasks and compared it to previous\nmethods. Results suggest that our method compares favorably with previous work.\nBesides, we present the efficiency of our approach for detecting wrongly\nclassified and adversarial examples.\n