Weannet enables parameter-efficient transfer through decaying lateral connections in progressive networks

Abstract Progressive Neural Networks (PNNs) transfer knowledge by connecting each new column to frozen earlier columns, but retained state and inference work grow across generations. I propose WeanNet, a bounded alternative in which one frozen parent supplies layer-wise features to a freshly initialized student while a scalar gain decays during training; the parent is then removed. The primary schedule decays to zero, making removal exactly lossless. Evaluation includes a controlled concept-drift task across entropy settings and three multilayer-perceptron architectures, a sequential-gravity CartPole stress test, and a five-seed Roblox obstacle course. On concept drift, WeanNet had higher mean drift-found accuracy than the architecturally matched PNN-Reset control at every tested entropy setting, although it was comparable to Reset rather than uniformly best. A tuned full PNN remained weak when the direct parent-logit pathway was enabled, but had substantially higher mean reward when that pathway was disabled, showing that its poor performance was configuration-dependent. CartPole favored Distillation and Regular, and Roblox showed overlapping seed variability, delimiting the generality of the concept-drift result. With a zero terminal gain, parent removal caused no return, logit, or action change and reduced deployed resident parameters. Positive terminal gains could improve attached performance but introduced a measurable pruning gap. These results support bounded temporary feature transfer, not universal performance superiority.