DFA-GNN Forward Training
🔗 Source: arXiv
DFA-GNN: FORWARD LEARNING OF GRAPH NEURAL NETWORKS BY DIRECT FEEDBACK ALIGNMENT
🚀 Technical Novelty
- Mechanism: Extends Direct Feedback Alignment (DFA) to GNNs by routing global output errors through fixed, topology-informed feedback links and introducing a pseudo-error generator that propagates residual supervision from labeled nodes to unlabeled ones.
- Nuance: Unlike prior non-BP methods that rely on data augmentation or layer-wise local losses, DFA-GNN directly projects supervision signals to hidden layers without negative samples or virtual nodes, maintaining gradient alignment while bypassing backpropagation’s weight-transport and update-locking bottlenecks.
💡 Yield
- Surpasses standard BP and state-of-the-art non-BP baselines across 10 benchmarks in accuracy, robustness to structural noise/attacks, and scalability to large graphs with reduced memory footprint.
- Provides a formal convergence proof for the proposed forward learning framework and demonstrates seamless portability across mainstream GNN architectures (GCN, GAT, GraphSAGE, etc.).
⚠️ Limitations
- Relies on fixed random feedback matrices whose alignment with forward weights may require careful initialization or tuning to maintain optimal gradient congruence.
- Slightly slower training convergence than BP due to non-exact gradient direction alignment, and primarily validated on semi-supervised node classification rather than broader graph tasks.