Explaining Machine Learning: Adding Interactivity to Develop Decision-making Visualization Expectations

Multi-layered Perceptrons, also widely known as Artificial Neural Networks, are one of the most used subsets of Machine Learning systems. They consist of a series of layered and interconnected Perceptrons, also referred to as Artificial Neurons, where each layer can perform different types of transformations on data. Because Multi-layered Perceptrons can consist of 10 to 30 layers, these transformations quickly become extremely complex and incomprehensible and these systems therefore become black boxes. This, coupled with the increasing presence of these systems in our lives, has led us to focus on the necessity of their humanunderstandability and transparency. As Information Visualization has proved to be one of the most effective ways of explaining datasets, our hypothesis is that it can be equally useful for understanding the processes of decision-making in Machine Learning systems. We approached this by developing an algorithmic data-driven visualisation of the most elementary processing element in these systems: the Perceptron. Our latest iteration, which received a user interface, has shown us to allow debugging, generating human expectation of the Perceptron's decision-making. This interdisciplinary and practice-based research focused on investigating the inner workings of a Perceptron and was developed through iterative cycles of research, design and implementation.