Robots are expected to perform smart services and to undertake various troublesome or difficult tasks in the place of humans. Since these human-scale tasks consist of a temporal sequence of events, robots need episodic memory to store and retrieve the sequences to perform the tasks autonomously in similar situations. As episodic memory, in this talk, we introduce a novel Deep adaptive resonance theory (ART) neural model and apply it to the task performance of robots. In addition, we propose Developmental Resonance Network (DRN) to overcome the limitation of Deep ART that requires the input data to be normalized between zero and one in advance. As an advanced version of DRN, adaptive DRN (A-DRN) is proposed to train the network without pre-tuning a vigilance parameter, which is the crucial hyperparameter on the network performance. Finally, we introduce Episodic Memory-DRN-MAP (EDM) to map multiple EM-DRNs using a new column map structure. EDM can directly learn multiple episodes from raw input data by EM-DRNs, which is the hybrid version of Deep ART and DRN. On top of that, each node in a MAP field comprises a column that indicates the specific label and has many cells mapping multiple episodes. EDM can be applied to recommender systems, e.g., the agent can recommend suitable recipes based on the currently existing ingredients and user preference. From the real robot experiments, we demonstrate the effectiveness of the proposed networks.
Gyeong-Moon Park received the B.S. degree in electronic and electrical engineering from Sungkyunkwan University, Suwon, Korea in 2014, and the M.S. and Ph.D. degrees in electrical engineering from KAIST, Daejeon, Korea in 2016 and 2019, respectively. Currently, he is working as a postdoctoral researcher at KAIST. His research interests include autonomous robots, online incremental learning, continual learning, and deep reinforcement learning. He is also working on deep learning for image-to-image translation and recommender systems. Dr. Park developed novel episodic memory networks for task intelligence of robots. He proposed Deep Adaptive Resonance Theory (Deep ART) as an episodic memory of robots, which can learn and recall the task episodes to perform tasks autonomously. To overcome the limitations of the ART networks, he proposed Developmental Resonance Network (DRN) and Adaptive DRN (A-DRN). DRN can learn raw input data directly without normalization, and on top of that, A-DRN does not need to pre-tune a vigilance parameter as hyperparameter. Besides, he proposed EM-DRN-MAP (EDM), the integrated memory combining Deep ART, DRN, and a new column map field. EDM can learn multiple episodes from the labeled data and can be applied to recommender systems.
Gyeong-Moon Park, Ph.D., Post-doctoral Researcher
Information & Electronics Research Institute, College of Engineering, KAIST
291 Daehak-ro, Yuseong-gu, Daejeon 34141, Republic of Korea
Tel: +82-42-350-8874, Mobile: +82-10-8840-7557