Keynote Speakers

Abstract: Deep Learning justifiably attracted the attention and interest of the scientific community and industry as well as of the wider society and even policy makers. However, the predominant architectures (from Convolutional Neural Networks to Transformers) are hyper-parametric models with weights/parameters being detached from the physical meaning of the object of modelling. They are, essentially, embedded functions of functions which do provide the power of deep learning; however, they are also the main reason of diminished transparency and difficulties in explaining and interpreting the decisions made by deep neural network classifiers. Some dub this “black box” approach. This makes problematic the use of such algorithms in high stake complex problems such as aviation, health, bailing from jail, etc. where the clear rationale for a particular decision is very important and the errors are very costly. This motivated researchers and regulators to focus efforts on the quest for “explainable” yet highly efficient models. Most of the solutions proposed in this direction so far are, however, post-hoc and only partially addressing the problem. At the same time, it is remarkable that humans learn in a principally different manner (by examples, using similarities) and not by fitting (hyper-) parametric models, and can easily perform the so-called “zero-shot learning”. Current deep learning is focused primarily on accuracy and overlooks explainability, the semantic meaning of the internal model representation, reasoning and decision making, and its link with the specific problem domain. Once trained, such models are inflexible to new knowledge. They cannot dynamically evolve their internal structure to start recognising new classes. They are good only for what they were originally trained for. The empirical results achieved by these types of methods according to Terry Sejnowski “should not be possible according to sample complexity in statistics and nonconvex optimization theory”. The challenge is to bring together the high levels of accuracy with the semantically meaningful and theoretically sound and provable solutions.

All these challenges and identified gaps require a dramatic paradigm shift and a radical new approach. In this talk, the speaker will present such a new approach towards the next generation of explainable-by-design deep learning. It is based on prototypes and uses kernel-like functions making it interpretable-by-design. It is dramatically easier to train and adapt without the need for complete re-training, can start learning from few training data samples, explore the data space, detect and learn from unseen data patterns. Indeed, the ability to detect the unseen and unexpected and start learning this new class/es in real time with no or very little supervision is critically important and is something that no currently existing classifier can offer. This method was applied to a range of applications including but not limited to remote sensing, autonomous driving, health and others.