WeanNet transfers learning through fading neural connections

A reinforcement-learning system called WeanNet can transfer useful features from a trained parent network to a newly initialized student, then gradually withdraw that support so the student can operate independently. In a controlled concept-drift experiment, WeanNet found newly correct actions more reliably than an otherwise matched progressive-network control that retained full-strength lateral connections at every tested entropy level. The result did not amount to universal superiority: learning from scratch matched or slightly outperformed WeanNet on several measures, and exploratory tests in CartPole and Roblox produced mixed results. Its clearest demonstrated benefit was bounded temporary transfer that allowed exact removal of the parent after training.

The method addresses a structural problem with progressive neural networks, or PNNs. A conventional PNN adds a new trainable column for each generation and connects it to all frozen earlier columns. This preserves access to previous representations, but the number of retained columns, evaluated parameters and lateral connections grows as the sequence continues. WeanNet instead keeps only one frozen parent and one student during a generation. At the corresponding hidden and output layers, the parent’s activations enter the student through trainable lateral matrices. The student is not instructed to reproduce the parent’s output distribution, as it would be under knowledge distillation; instead, task rewards determine how useful each inherited feature becomes.

A shared scalar gain controls the influence of those parent features. In the principal schedule, it declines linearly from 1 to zero during the first 300 of 600 training batches. As support fades, the student must encode useful behavior in its own weights rather than continuing to depend on the previous network. When the gain reaches zero, every contribution from the parent is mathematically zero, so deleting the parent cannot alter returns, logits or selected actions. This property was verified in the concept-drift experiment and across all 40 post-shift CartPole seed-generation comparisons. In the 24-64-64-8 concept-drift network, pruning reduced resident parameters from 17,232 to 6,280 per agent, a 63.6% reduction.

The central confirmatory comparison used PNN-Reset, a control with the same one-parent architecture, lateral matrices and injected activations as WeanNet, except that its gain remained fixed at 1 throughout training. The synthetic environment contained 64 states and eight actions, with a new correct action assigned to 10% of states in every generation after the first. For a changed state, the new action earned +1, the previously correct action remained a decoy worth +0.5, and other actions earned -0.1. Each episode contained one decision, isolating the ability to escape an outdated action from longer-horizon credit assignment. Results were averaged across eight agents on each of 10 environment layouts, with variability summarized across the 10 layout-level averages.

WeanNet’s drift-found accuracy exceeded that of PNN-Reset at all four tested policy-entropy coefficients. The gaps were 0.213 at entropy 0, 0.213 at 0.05, 0.215 at 0.1 and 0.190 at 0.2. It also accumulated less “drift-trap” mass, 0.248 versus 0.292, meaning that PNN-Reset more often continued choosing the old, locally rewarding action. The pattern appeared across three multilayer-perceptron structures: 32-32, 64-64 and 64-64-64 hidden units. Nevertheless, the mechanism was not cleanly isolated. WeanNet also had higher core-state accuracy than PNN-Reset in the primary comparison, contrary to a preregistered expectation that any advantage should be concentrated on changed states. One proposed explanation is that decaying support gives the active column more practice operating independently before it becomes the next parent, but the experiments did not directly test that generation-handoff mechanism.

The strongest caution comes from the no-transfer Reset baseline. At entropy 0, Reset achieved the highest overall reward, 0.584, narrowly ahead of WeanNet’s 0.578, and it also had higher drift-found accuracy, 0.535 versus 0.488. WeanNet’s core accuracy of 0.567 was close to Reset’s 0.558 and Warm-Start’s 0.561, while Regular had the highest core accuracy at 0.599. Across the three architecture variants, WeanNet’s mean rewards were also close to those of Reset. Because the task was small and relearning fit within the training budget, starting without inherited preferences remained competitive. The findings therefore support an advantage over full-strength one-parent transfer in this setting, not an advantage over learning from scratch or across all transfer methods.

Schedule sensitivity revealed a deployment trade-off. Raising the terminal gain floor from zero to 0.05 increased attached core accuracy from 0.567 to 0.683 and attached reward from 0.578 to 0.683, while drift-found accuracy changed little, from 0.488 to 0.492. But a positive floor meant that deleting the parent changed the model: pruning at 0.05 cost an average 0.026 in core accuracy and 0.029 in reward. A floor of 0.1 reduced drift-found accuracy to 0.421 and created a larger pruning gap. Lower initial gains also performed better in the tested range, with rewards of 0.725, 0.679 and 0.578 for initial gains of 0.25, 0.5 and 1. Because 0.25 was the lowest value examined, it cannot be treated as a universal optimum. These schedule analyses used the same evaluation layouts as the main experiment, so they characterize sensitivity rather than held-out schedule selection.

The results also showed that the direct parent-to-output pathway was a major source of configuration dependence. A full PNN performed poorly with that pathway enabled, reaching mean reward of 0.191 in the primary concept-drift test, but its reward rose to 0.665 when the direct logit connection was disabled. WeanNet and PNN-Reset likewise scored 0.578 and 0.449 with the pathway on, compared with 0.713 and 0.653 when it was off. These findings do not support a claim that growing PNNs inevitably collapse. The comparison also involved different tuning procedures: the full PNN’s learning rate and lateral initialization were selected on separate validation layouts, whereas WeanNet’s primary decay schedule was held fixed. Its selected PNN settings were both at the lower boundaries of the search grids, so they represent the best observed configuration rather than proof of complete optimization.

Exploratory stress tests placed limits on broader claims. In five CartPole generations with gravity changing from 9.8 to 14.7, 4.9, 19.6 and back to 9.8, Distillation and Regular achieved the highest mean post-shift returns, 476.12 and 472.99. WeanNet reached 394.75, compared with 395.86 for Reset and 401.86 for PNN-Reset. Exploratory paired confidence intervals did not resolve differences among these three methods, but their width also prevented a claim of practical equivalence. In a five-seed Roblox obstacle-course experiment, final training-reward means overlapped across leading methods; WeanNet recorded 0.2838 ± 0.0403. The stored Roblox results measured reward rather than direct visits to the newly opened region, so they could not establish that any agent escaped the local optimum. Overall, the evidence supports decaying lateral transfer as a useful mechanism in controlled concept drift, while showing that its performance benefits depend on the task, architecture and training configuration.

Subject of Research: A bounded reinforcement-learning architecture that transfers parent-network features through decaying lateral connections

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

Article References: Vorachart, M. (2026). Weannet enables parameter-efficient transfer through decaying lateral connections in progressive networks. Discover Artificial Intelligence, 6(1), Article 1564. https://doi.org/10.1007/s44163-026-02490-8

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02490-8

Keywords: reinforcement learning, continual learning, concept drift, progressive neural networks, knowledge transfer, lateral connections, parent pruning, policy adaptation, CatPole, Roblox, Weannet, enables