In this paper, we study the task of facial expression recognition under\nstrong occlusion. We are particularly interested in cases where 50% of the face\nis occluded, e.g. when the subject wears a Virtual Reality (VR) headset. While\nprevious studies show that pre-training convolutional neural networks (CNNs) on\nfully-visible (non-occluded) faces improves the accuracy, we propose to employ\nknowledge distillation to achieve further improvements. First of all, we employ\nthe classic teacher-student training strategy, in which the teacher is a CNN\ntrained on fully-visible faces and the student is a CNN trained on occluded\nfaces. Second of all, we propose a new approach for knowledge distillation\nbased on triplet loss. During training, the goal is to reduce the distance\nbetween an anchor embedding, produced by a student CNN that takes occluded\nfaces as input, and a positive embedding (from the same class as the anchor),\nproduced by a teacher CNN trained on fully-visible faces, so that it becomes\nsmaller than the distance between the anchor and a negative embedding (from a\ndifferent class than the anchor), produced by the student CNN. Third of all, we\npropose to combine the distilled embeddings obtained through the classic\nteacher-student strategy and our novel teacher-student strategy based on\ntriplet loss into a single embedding vector. We conduct experiments on two\nbenchmarks, FER+ and AffectNet, with two CNN architectures, VGG-f and VGG-face,\nshowing that knowledge distillation can bring significant improvements over the\nstate-of-the-art methods designed for occluded faces in the VR setting.\n