DOCTORAL CANDIDATES

Description

Effective communication is fundamental to progress in our world, enabling the exchange of information and interaction between humans and machines. The rising interest in Artificial Intelligence (AI) underscores the importance of fostering a positive human-machine relationship, with the Tactile Internet emerging as a key facilitator in this regard.

The Tactile Internet aims to augment human capabilities rather than replace humans with technology, focusing on human-to-machine/robot (H2M/R) interactions. Defined as “a network, or a network of networks, for remotely accessing, perceiving, manipulating, or controlling real and virtual objects or processes in perceived real-time," the Tactile Internet represents a significant leap forward in human-machine synergy [ [1]. A critical component of advancing the Tactile Internet is the implementation of continual learning. Continual learning is defined as “an adaptive learning process that incorporates real-time feedback and interaction, enabling systems to acquire, process, and retain new information without forgetting previously learned knowledge”. In other words, when a pre-trained model encounters new tasks it should be able to deduce the presence of new tasks solely through input characteristics and try to keep acquired knowledge from the previous task. Moreover, the pre-trained model efforts to keep prior knowledge by regularization on parameters, activation, hyperparameters, or parameter isolation whether fixed or dynamic approach or memory-based approaches [2]. Traditional artificial neural networks (ANNs) often struggle with this, experiencing catastrophic forgetting which is learning sequentially on multiple tasks and tends to forget previously learned information upon acquiring new tasks. This issue arises because the weight adjustments made for learning a new task can significantly interfere with and overwrite the weights associated with previous tasks [3]. In the Tactile Internet, we deal with real-time data streams associated with specific tasks. Utilizing continual learning enables us to effectively analyze this dynamic data, which changes the distribution of input data over time. In the study of continual learning, scenarios are defined based on variations between the input distributions and target label distributions of previous and new tasks [4].

This research project aims to propose a novel and improved continual learning method within the Tactile Internet framework, enhancing AI systems' ability to adapt to changing environments and improve human interactions. By integrating continual learning, the project seeks to create a symbiotic relationship between humans and machines, where both parties learn from each other and collectively enhance their capabilities. The ultimate goal is to achieve a harmonious partnership, with technology augmenting human potential and driving innovation and progress.

[1] Fettweis,Gerhard P., "The Tactile Internet Applications and Challenges," IEEE Vehicular Technology Magazine, pp. 64-70, 2014.

[2] Wickramasinghe, Buddhi and Saha, Gobinda and Roy, Kaushik, "Continual Learning: A Review of Techniques, Challenges and Future Directions," IEEE Transactions on Neural Networks and Learning Systems, vol. 35, pp. 123-140, 2024.

[3] Kirkpatrick, James and Pascanu, Razvan and Rabinowitz, Neil and Veness, Joel and Desjardins, Guillaume and Rusu, Andrei A. and Milan, Kieran and Quan, John and Ramalho, Tiago and Grabska-Barwinska, Agnieszka and Hassabis, Demis and Clopath, Claudia and Ku, "Overcoming Catastrophic Forgetting in Neural Networks," Proceedings of the National Academy of Sciences, vol. 114, no. 13, pp. 3521-3526, 2017.

[4] Konstantinos Antonakoglou, Xiao Xu, Eckehard Steinbach, Toktam Mahmoodi, Mischa Dohler, "Toward Haptic Communications Over the 5G Tactile Internet," IEEE Communications Surveys \& Tutorials, pp. 3034-3039, 2018.