Automatic Model Adaption Method based on Few-Shot Incremental Learning for $\text{IoT}$ Applications
Dequn Kong, Xiaotao Li, Wai Chen · 2022 IEEE/CIC International Conference on Communications in China (ICCC) · 2022
The dynamic Internet-of-Things application scenarios can lead to model expiration. Machine learning models need to update adaptively and efficiently. In most cases, models are expected to learn new classes from consecutive tasks. Attempts to develop an incremental learning system have been impeded by a chronic problem called “catastrophic forgetting”. This problem will become worse when the number of samples of the new classes is low. To alleviate the effect of catastrophic forgetting, it's essential to retain the previously learned knowledge and accommodate new knowledge over time. Inspired by the dual-coding theory, we propose the SE, a novel computationally efficient model adaption framework with a cooperative model architecture consisting of a sampling model (sampler) and an incremental classifier (extractor). With these two components (the sampler, the extractor), knowledge can be easily accumulated, and incremental classifier can be achieved. Extensive experiments have been carried out to demonstrate the effectiveness and efficiency of the proposed SE approach.