ICAISC 2020

Artificial Intelligence has great impact on every aspect of technology, including neurotechnologies used for human augmentation. In recent years progress in methods of brain activity measurement, analysis of neuroimaging and electrophysiological data, and understanding of brain processes, opens new areas for transdisciplinary applications. Identifying patterns of EEG/MEG, ECoG or fMRI signals that serve as "fingerprints" of high subnetwork activity allows for many applications: linking brain activity with thoughts, intentions, emotions and other mental states, objective diagnostic methods in neuropsychiatry, reliable brain-computer interfaces (BCI), optimization of brain processes through neurofeedback, therapeutic interventions using neuromodulation, neurorehabilitation based on direct brain stimulation combined with behavioral procedures. Some commercial applications for treating epilepsy, major depression and other mental problems are already on the market.
Although I will focus on technical aspects of brain fingerprinting it is worth to reflect how neurocognitive technologies will enable human-computer interaction, and in an unprecedented way will change the very nature of people, their social interactions and coupling with physical environment.

Human beings reflect their internal states in many ways in their physiological signals, from skin conductivity, heart rate, pupil dilation, brain signals and behavioural measures. Many of these can be collected unobtrusively. The kinds of internal states we have investigated include stress, depression, emotion veracity, and doubt. We have shown in a number of such areas that physiological signals recorded from a human observer can be used to predict the ground truth in the observed data better than the same human beings themselves can do. That is, by the use of appropriately cross-validated machine learning training, we can access implicit knowledge within the human participants, which is not available to their consciousness.

In the last few years, interpretability of classification models has become a very active area of research. Both ACM and IEEE initiated new interdisciplinary conferences where fairness, accountability and transparency of "black-box" algorithms is the main topic. Suddenly, computer programs are being evaluated from moral and political point of view.

In this talk, I will discuss a couple of recent controversies in this area. First, I will talk briefly about a supposed racial bias in a COMPAS system, widely used in US courts. Second, I will discuss the concept of algorithm interpretability in a more specific legal context. In 2018 EU introduced General Data Protection Regulation with a Right to Explanation for people subjected to automated decision making. The Regulation itself is very brief on what such a right might imply. I will attempt to explain what the Right to Explanation may involve. I then will argue that this right would be very difficult to implement due to technical challenges. I also maintain that the Right to Explanation may not be needed and sometimes may even be harmful.

Learning from imbalanced data is considered one of the vital challenges in contemporary machine learning. Despite more than three decades of research, the problem of handling skewed distributions is still as important as ever, with new challenges emerging on regular basis. This talk will give an overview of the imbalanced learning domain, focusing on contemporary challenging scenarios and recent developments. Special attention will be given to data-level difficulties and understanding minority classes, multi-class imbalanced problems, and data streams with dynamically evolving classes. The talk will discuss various resampling methods, low-dimensional embeddings, algorithm-level modifications, and ensemble learning approaches that were recently proposed to efficiently handle such challenging scenarios.