In the development of real-world artificial intelligence systems, the capability of incrementally learning new data is a crucial challenge. Few-Shot Class-Incremental Learning (FSCIL) combines class incremental learning and few-shot learning. Class incremental learning (CIL) is one of the incremental learning concepts that train the learning model to learn the new class in a large number of tasks without excessive computational and memory growth. CIL involves forward transfer of knowledge and backward transfer of new data.
The main problem in CIL is catastrophic forgetting, and it is the way of learning model that abruptly forgets the previously learned data upon learning new data. CIL utilizes knowledge distillation to store the old class in external memory. Few-shot learning is the learning technique in which the model learns the information with a limited number of data and reduces data collection effort and computational cost. FSCIL is the specific type of incremental learning in which the classifier learns a new class using very few training data without forgetting the previously learned data. The main aim of the FSCIL is to avoid catastrophic forgetting and prevent overfitting to new classes.FSCIL is applied in many tasks such as intrusion detection, object detection, image recognition, image classification, and so on.
• Few-shot class-incremental learning (FSCIL) encompasses two challenging problems such as few-shot learning and incremental learning aims to learn new categories with only a few data samples continuously.
• FSCIL is a recent approach for recognizing few-shot new classes without forgetting old classes, and its main significance is to address the few-shot inputs in incremental learning.
• Nevertheless, FSCIL faces some challenges involving over-fitting problems through few-shot training examples.
• The main challenge of FSCIL is the scarcity in the data of new classes, which not only causes severe overfitting but also exacerbates the catastrophic forgetting problem of old classes.
