Chapter 6 Conclusion and Future Work In this concluding chapter, we revisit the research questions posed at the start of this thesis and summarise the contributions we made to address them. Furthermore, we reflect on the impact and limitations of this thesis, and outline potential directions for future work. 6.1 Conclusions Our first research question asks how the density estimation of probabilistic models, in particular Variational Autoencoders (VAEs) (Kingma and Welling, 2013; Rezende et al., 2014), can help with anomaly detection. In Chapter 3, we present a framework to obtain anomaly scores for new data points, by first training a VAE on normal data and then using its density estimation to provide a value of how likely new data points fit the distribution of normal data. Less likely data points can then be detected as anomalies, by defining a threshold on this anomaly score. We test this anomaly framework on a simple benchmark dataset, as well as in an industrial setting for fault detection in 3D-printed products. Results show some success, implying that this is indeed a valid approach for anomaly detection. However, we also observe that in certain settings the VAE fails to capture the normal data well enough, leading to weaker anomaly detection results. This becomes particularly obvious in an additional use case of lung cancer detection in 3D images of lung nodules, where our results show that our models fail to
RkJQdWJsaXNoZXIy MjY0ODMw