| Continual learning requires neural networks to acquire new tasks sequentially without catastrophic forgetting. This study presents a systematic empirical investigation of six neural architectures --- Columnar Independent Networks (CIN), AutoEncoder Classifier, BasicConvNet, ResNet18, MobileNetV2, and DenseNet121---under a true progressive learning protocol (sequential training without resetting weights, using a multi-head output design). We extend conventional accuracy metrics to include standard continual learning measures: average accuracy (ACC), backward transfer (BWT), forward transfer (FWT), and forgetting. We demonstrate that while traditional architectures achieve superior single-task performance, CIN provides perfect task isolation (zero forgetting) at the cost of linear parameter growth. Moreover, we compare against established continual learning algorithms (EWC, GEM), showing that architectural choices can be as effective as algorithmic ones. Our multi-dimensional evaluation framework provides actionable insights for architecture selection in continual learning scenarios. Code is available at: https://github.com/mab85/Analysis-for-Multi-Task-ContinualLearning.git |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.