IEEE INTERNATIONAL CONFERENCE ON SYSTEMS, MAN, AND CYBERNETICS 2020
Special Session: Human-in-the-loop Machine Learning and Its Applications
Toronto, Canada, OCT 11-14, 2020
We will be closely following the update of COVID-19 and the organizing of SMC 2020.
- Deadline for Submission: May 15, 2020 (extended)
- Acceptance Notification: May 30, 2020
- Conference Dates: Oct 11-14, 2020.
Please visit http://smc2020.org/submission/ and submit the paper on the SMC2020 system and select "Special Session Paper" . Enter the code "qfi15" in the submission page.
The selected papers are encouraged to submit their extended versions to the Topical Collection of the Neural Computing and Applications (Springer, IF=4.664 (2018)).
Human-in-the-Loop (HIL) means including human feedback into the training loop of the machine learning models in order to facilitate the following requirements:
1) to improve the quality of training and reduce/prevent the error of the model. When the testing error is larger than a certain threshold, the HIL learning model is able to obtain the new data-points from the users in an interactive way. In some situations, a large error produced by the model should be avoided. For instance, reinforcement learning alone is not sufficient to achieve safety if there exists an exploration policy in robot manipulation, by which some unexpected actions may be generated. In such scenarios, the data-points from the human guidance are crucial during both robot’s safe execution as well as model optimization.
2) to incorporate the human user labelling to improve the pre-trained models. During the training of the state-of-the-art models, the quality of the training data-sets is extremely important. One solution to actively incorporate more data is optimizing the models by including the human users’ feedback (e.g. rewards in RL) or new data-points (e.g. supervised learning) to adapt the pre-trained models in different environments.
In the aforementioned requirements, humans are involved in the training process of the algorithms by continuously optimizing the model’s parameters, feeding the data or even adjusting the model itself by meta-learning. From the perspective of algorithm design, a key problem to design a proper training with a human in the loop is how to leverage both active learning from a human and the optimization of the models. In other words, how can we design a proper query strategy depending on different applications and scenarios?
When properly implemented, the HIL is suitable to be applied in real-world applications where the data is sparse. The active learning mechanism built in the model can be helpful which could seek the human’s help in a form of supervised or reinforcement learning. In this way, proper designs of interactive displays, machines and robots could be of help to obtain the human’s inputs. They are related to designs of HCI, UX/UI, etc, and related to how we can efficiently utilize the human expertise to reduce the exponential search space. In this special session, the designs and experiments can also be discussed to evaluate the effectiveness of human-in-the-loop applications.
Specifically, following the success and discussions in the IEEE Symposium in Domestic Robotics in 2019, we are particularly encouraging robotic applications and their experimental deployment using HIL algorithms. We believe that the HIL algorithms will be an effective method to make robotic platforms more adaptive and safer to interact with. The workshop will offer the opportunity for researchers and practitioners in the diverse field where human reinforcement feedback would have a positive impact on the training processes. The inclusion of HIL would allow robots and machine learning models to use both internal and external feedback to speed up the learning process and also improve its performance. In many ways this could allow the models to learn through their own self-reflection as well as the external input from a human.
Topics of interest include, but are not limited to:
Human Guided Reinforcement Learning
Human-robot Social Interaction
Dialogue Systems with Human-in-the-loop
Interpretable Machine Learning with Human-in-the-loop
Active Learning and Continuous Learning
Learning by Demonstration
Human Factors in HCI/HRI
Joni Zhong, Nottingham Trent University, UK
Mark Elshaw, Coventry University, UK
Yanan Li, University of Sussex, UK
Stefan Wermter, University of Hamburg, Germany
Xiaofeng Liu, Hohai University, China