Fault localisation (FL) for DNNs is a rapidly evolving area of DL testing [1-5]. The decision logic of traditional software systems is encoded in their source code. Correspondingly, fault localisation for such systems consists of identifying the parts of code that are most likely responsible for the encountered misbehaviour. Unlike traditional software systems, however, the decision logic of DL systems depends on many components such as the model structure, selected hyper-parameters, training dataset, and the framework used to perform the training process. Moreover, DL systems are stochastic in nature, as a retraining with the exactly same parameters might lead to a slightly different final model and performance. These distinctive characteristics make the mapping of a misbehaviour (e.g., poor classification accuracy) to a specific fault type a highly challenging task. Existing state-of-the-art works [1, 2, 4, 6, 7] that focus on the problem of fault localisation for DL systems were shown to be adequate for this task when evaluated on different sets of real-world problems extracted from StackOverflow and GitHub platforms or were deemed useful by developers in the process of fault localisation and fixing [5].