KEYNOTE SPEECH 1.  5/25(수) 10:00 ~ 10:30 (30")

Dr. Wojciech Samek

Fraunhofer Heinrich Hertz Institute

Towards Actionable Explainable AI

Explainable Artificial Intelligence (XAI) is an emerging research field bringing transparency to highly complex and opaque machine learning (ML) models. Despite the development of a multitude of methods to explain the decisions of black-box classifiers in recent years, these tools typically provide rather limited information (e.g., simple attribution map) and are seldomly used beyond visualization purposes. Thus, decisions are explained and problems may be discovered, but the obtained insights are rarely applied to actually achieve more trustworthy, fairer, or simply better performing models. This talk presents recent developments to close this gap and make explanations semantically richer, more easy to understand, and in conclusion more actionable. First, we present Concept Relevance Propagation (CRP), an extension to the Layer-wise Relevance Propagation framework allowing to increase the expressive power of local XAI by connecting explanations to human-understandable concepts. Second, we demonstrate that explanations can be used beyond mere visualization purposes and can improve properties such as model generalization or efficiency, among others.

KEYNOTE SPEECH 2.  5/26(목) 10:00 ~ 10:30 (30")

Prof. Geoffrey Ye Li

Imperial College London

From Conventional to Semantic Communications based on Deep Learning

To transmit text messages, speeches, or pictures, we usually convert them into a symbol sequence and transmit the symbols in a conventional communication system, which is designed based on the block structure with coding, decoding, modulation, demodulation, etc. It has been demonstrated recently that deep learning (DL) has great potentials to break the bottleneck of the block-based communication system. In this talk, we first provide our recent endeavors in developing end-to-end (E2E) communications, which combine all blocks at the transmitter by a neural network and those at the receiver by another neural network. Even if deep learning based E2E communication systems have a potential to outperform the conventional block-based communication systems in terms of performance and complexity, their spectrum efficiency is still limited by Shannon capacity since they essentially transmit bits or symbols. Semantic communication systems transmit and recover the desired meaning of the transmitted content (for example, a text message or a picture) directly and can significantly improve transmission efficiency. We will present our initial results on semantic communications.