Abstract
In this paper, we present a novel incremental learning technique to solve the catastrophic forgetting problem observed in the CNN architectures. We used a progressive deep neural network to incrementally learn new classes while keeping the performance of the network unchanged on old classes. The incremental training requires us to train the network only for new classes and fine-tune the final fully connected layer, without needing to train the entire network again, which significantly reduces the training time. We evaluate the proposed architecture extensively on image classification task using Fashion MNIST, CIFAR-100 and ImageNet-1000 datasets. Experimental results show that the proposed network architecture not only alleviates catastrophic forgetting but can also leverages prior knowledge via lateral connections to previously learned classes and their features. In addition, the proposed scheme is easily scalable and does not require structural changes on the network trained on the old task, which are highly required properties in embedded systems.
Highlights
Network for Incremental Learning.“Tell me and I forget, teach me and I may remember, involve me and I learn.” (Benjamin Franklin)
In addition to comparing the proposed progressive incremental learning (PIL) method with joint training (JT), we compared PIL with a network structure that trains an independent classifiers on each subset of incremental classes, without any lateral connections
We call the implementation without lateral connections PIL with no lateral connections (PIL-NLC)
Summary
Network for Incremental Learning.“Tell me and I forget, teach me and I may remember, involve me and I learn.” (Benjamin Franklin). Natural learning systems are inherently incremental [1] where the learning of new knowledge is continued forever, ideally. Despite the staggering recognition performance of convolutional neural networks (CNN), one area where the CNNs struggle is the catastrophic forgetting problem [2]. The degradation in the performance of a CNN network on old classes (or tasks) when a pre-trained CNN is further trained on new classes, is referred to as a catastrophic learning problem. Some researchers have termed this problem as a domain expansion problem [3]. This happens since the weights in the network that are important for old classes are updated to accommodate the new classes
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
