Interictal discharges (IEDs) in EEG recordings are important signatures of epilepsy as their presence is strongly associated with an increased risk of seizures. IEDs are relatively short-duration events (typically 70–250 ms) that can be viewed as stochastic anomalies in such recordings. Currently, visual analysis of the EEG by clinical experts is the gold standard. This process, however, is time-consuming, error prone, and associated with a long learning period. Automatizing the detection of IEDs has the potential to significantly reduce review time, and may serve to complement the visual analysis. Supervised deep learning methods have shown potential for this purpose, but the scarceness of annotated data has limited their performance, which motivates to explore unsupervised and semi-supervised approaches, that do not require (extensive) expert annotations. We trained different unsupervised deep learning models, Autoencoders (AE) and Variational Autoencoders (VAE) for anomaly (IED) detection in these recordings. These models are dimensionality reduction based approaches, that can compress the data to lower dimensional representations, learning the notion of normality within data and reconstruct samples accordingly. Our data set comprised 203 clinical EEGs, 115 from patients with epilepsy, that contained IEDs, and 88 normal EEGs. Performance was assessed qualitatively through visual analysis of reconstructed samples and quantified as Area Under the Curve (AUC), sensitivity and specificity. The best performance was obtained using a semi-supervised approach, allowing the detection of IEDs with a sensitivity of 81.9% and specificity of 91.7%. Our work shows that unsupervised approaches and other approaches with limited supervision perform satisfactorily and have the potential to assist visual assessment of interictal discharges in epilepsy diagnostics. [Display omitted] • Interictal epileptiform discharges (IEDs) are associated with seizure risk, but detecting them is labor-intensive. Automation can help experts' visual analysis. • The need for more labeled data limits supervised methods. Learning from limited or unlabeled data shows promise. • Our unsupervised approach provides concise EEG signal representations and enables valuable insights into the location of the abnormalities. • Incorporating labels and learning constraints guides robust semi-supervised models, competing with fully supervised ones. [ABSTRACT FROM AUTHOR]