ICMLA AML-IoT FLAME 2021 : IEEE Int'l Conf. on Machine Learning and Applications; Special Session on Advanced Machine Learning and Applications: Federated Learning and Meta-Learning,
Call For Papers
During the submission using:
Please choose "Special Session 4: Advanced Machine Learning and Applications: Federated Learning and Meta-Learning" for your submission.
Artificial intelligence (AI) and machine learning (ML) are key enabling technologies for many Internet of Things (IoT) applications and meta-learning. However, the collection and processing of data for AI and ML is very challenging in the IoT domain, even learning from data is critical in meta-learning. The most recent challenges that researchers tackle with are as follows:
1. There are usually a large number of low-powered sensors deployed in large geographical areas with possibly intermittent network connectivity.
2. The sensors and their collected data may be owned by different users or organizations, which can bring further obstacles to data collection due to privacy concerns and noisy labels provided by different users.
3. The successful application of AI/ML approaches in such scenarios with noisy and decentralized data is difficult.
4. The amount of collected data that can be used for training AI/ML models is usually proportional to the number of users in the system, but the system may not be able to attract many users without a well-trained AI/ML model. It is challenging to solve this dilemma.
5. One of the factors that is often neglected in data generation/collection is lack of sufficient sensing information that describe the real-world system. Further, there are a limited number of instances for each sensing category which causes inaccurate learning models. In these two cases, traditional machine learning algorithms have difficulties to understand the unseen and limited number of categories of samples.
6. Due to the increase of unlabeled datasets, supervised learning cannot make use of unlabeled data.
7. Inefficient labor work of labeling data makes generated data useless.
8. Low-performance classification due to unlabeled and unseen classes, that is the foundation of
8. While traditional ML approaches rely on labeled and observable datasets, there is a crucial need to develop learning algorithms that are capable of training the model using limited data to learn the concept of domain and problem. It means that machine learning algorithms cannot perceive the problem environment before the training process. However, human beings learn the problem by definition and the idea behind that. Hence, there is a need to revisit existing machine learning techniques.
10. Another yet important emerging problem is the high dimensionality problem which leads to the curse of dimensionality (CoD). Although there have been a large number of machine learning algorithms to deal with this problem, they still cannot guarantee that with the emerging data type their solution still works, so we need to have advanced data analytics algorithms to solve this problem.
This workshop focuses on how to address the above and other unique challenges of applying AI/ML in IoT and meta-learning systems. It invites researchers and practitioners to submit papers describing original work or experiences related to the entire lifecycle of an IoT powered by AI and ML, specifically meta-learning systems. Topics for this workshop include, but are not limited to:
· Techniques for making use of data collected by geographically dispersed sensors to provide useful services through AI/ML.
· Techniques for sharing data and training AI/ML models while preserving user sensitive information.
· Techniques for dealing with noisy data and labels.
· Techniques for reducing human effort in data labeling (such as active learning).
· Techniques for evolving from a new system that is initially trained with only a small amount of data.
· Techniques for automating the feature selection and hyperparameter tuning in the learning process and expediting the training with dynamic input datasets.
· Automated learning
· Efficient data analytics
· Distributed learning
· Federated learning and its applications
· Efficient learning on IoT devices
· Collaborative learning
Papers focusing on the above and other related topics are welcome. Specifically, federated learning and meta-learning techniques for IoT systems are the two main pivotal topics of this workshop. The uniqueness of our workshop is that it focuses on realistic problems that exist throughout the lifecycle of AI/ML-powered IoT and meta-learning systems, such as those mentioned above. Many of these problems have not been well addressed in existing conferences or workshops. For example, most existing work only addresses how to develop good AI/ML models while assuming that a centralized dataset already exists. We hope to bridge this gap with our workshop. Also, most studies assume there exists a sufficient training dataset, while in real-world scenarios there are many limitations on available datasets.