What is it about?

Consider a household robot using a Deep Neural Network (DNN) model to identify multiple objects. Changes in user requirements and environment, such as from daylight to nightlight, may compromise the robot’s recognition ability due to domain shift problems caused by changing data patterns. To overcome the performance degradation, existing DNN models in IoT need to retrain from scratch to adapt to domain shifts incrementally. To fill this gap, incremental learning is needed to enable DNN models to perform incrementally with domain shifts rather than being trained from scratch each time new data is collected. Taking inspiration from the neural-reuse principle of the human brain, we propose the algorithm E-DomainIL to enable IoTs to adapt incrementally to changing requirements and environments, just like humans.

Featured Image

Why is it important?

E-DomainIL achieves outstanding performance. Compared with the existing incremental learning methods, E-DomainIL has two advantages: more memory-efficient and requires less manual help; therefore it is more suitable to be applied to IoT systems.

Perspectives

The traditional deep neural network only imitates the most basic mechanism of the human brain, and many hyper mechanisms of brains need to be learned by machines. I believe this is a step towards more general artificial intelligence. I have confidence in this work as the results are robust and support our hypothesis. Our code is stable and easy to perform, and we are doing a lot of follow-up research at both algorithm and application levels.

Yuqing Zhao
Hong Kong Polytechnic University

Read the Original

This page is a summary of: Memory-Efficient Domain Incremental Learning for Internet of Things, November 2022, ACM (Association for Computing Machinery),
DOI: 10.1145/3560905.3568436.
You can read the full text:

Read

Resources

Contributors

The following have contributed to this page