In this paper, we address the problem of weakly-supervised video anomaly detection, in which given video-level labels for training, we aim to identify in test videos, the snippets containing abnormal events. Although current methods based on multiple instance learning (MIL) show effective detection performance, they ignore important video temporal dependencies. Also, the number of abnormal snippets can vary per anomaly video, which complicates the training process of MIL-based methods because they tend to focus on the most abnormal snippet -- this can cause it to mistakenly select a normal snippet instead of an abnormal snippet, and also to fail to select all abnormal snippets available. We propose a novel method, named Multi-scale Temporal Network trained with top-K Contrastive Multiple Instance Learning (MTN-KMIL), to address the issues above. The main contributions of MTN-KMIL are: 1) a novel synthesis of a pyramid of dilated convolutions and a self-attention mechanism, with the former capturing the multi-scale short-range temporal dependencies between snippets and the latter capturing long-range temporal dependencies; and 2) a novel contrastive MIL learning method that enforces large margins between the top-K normal and abnormal video snippets at the feature representation level and anomaly score level, resulting in accurate anomaly discrimination. Extensive experiments show that our method outperforms several state-of-the-art methods by a large margin on three benchmark data sets (ShanghaiTech, UCF-Crime and XD-Violence). The code is available at https://github.com/tianyu0207/MTN-KMIL