The next Doctoral Research Seminar is titled "Heideggerian AI: Don’t Program the Robot, Let It Learn" by Byoung-Tak Zhang, Professor of Computer Science, Director of the AI Institute (AIIS), Seoul National University
We discuss the role of robots (as embodied cognitive agents) for the advancement of AI (human-level AI) and the AI’s lessons for the advancement of robotics (building really smart robots!). Regarding the latter part, the suggestion is “Don’t try to program the robot, let it just learn itself”. Heideggerian epistemology has shown that human intelligence cannot be fully formalized and, thus, cannot be explicitly programmed by primitives. This explains why the classical AI, i.e. rule-based programming approach was not very successful. The Heideggerian existentialism suggests that, to achieve human-level intelligence, the agent should exist in the world (“being-in-the-world”) and cope with the world to learn the know-how. This also explains the superiority of the deep learning AI approaches to the previous, explicit programming approach: deep learning doesn’t pre-analyze situations in terms of a fixed set of possibly relevant features, it just experiences a large number of raw situations to learn holistically the implicit relevance. However, current learning approaches are still limited since they are blinded to (a la Heidegger) the unobservable situations that might be relevant (partial observability). This limitation can be overcome if we embody the agent with sensors and actuators. The embodied agent (cognitive robots) can learn fully autonomously even in partially observable environments by correcting its own actions from the perceptual self-feedbacks.