After learning to recognise familiar objects, people can easily recognise new ones, but machines cannot. Current state-of-the-art techniques for teaching machines to identify collections of items are based on deep neural networks (DNN). The problem is that any alteration to the DNN weights that were first used to categorise a group of objects might significantly reduce the network’s capacity to perform those original classifications; this phenomenon is known as catastrophic forgetting (CF). An architecture for continual learning (CL) that can handle CF is presented in this study. The architecture consists of two main components: (i) the modular dynamic classification (MDC) component and (ii) the feature extraction component, which is built on standard backbones. The first consists of several sub-networks that progressively form a tree-like structure. As it learns over time, it rearranges itself such that each sub-network may function on its own. The MDC is based on binary classifiers that are customized on the fly and may vary depending on the object. The second is the focus of the current study, where several backbones were tested and evaluated, while always maintaining the same MDC. The best results were achieved when applying the Swin Transformer as a backbone, and tests on the CORe50 dataset showed accuracy results of 91.0%.