Research map / Architectures and efficiency / RNNs and normalization
Multi-View Incremental Learning with Structured Hebbian Plasticity for Enhanced Fusion Efficiency
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MVIL, a bio-inspired multi-view incremental graph learning framework with structured Hebbian plasticity and synaptic partition learning, evaluated on semi-supervised node classification over six multi-view graph datasets.
Problem
Multi-view learning assumes all views are available in advance; when views arrive sequentially, models forget old knowledge and fail to fuse views at a fine-grained level. The paper targets multi-view incremental learning where each task adds one new view while all classes remain.
Approach
A GCN-based streaming model whose second-layer weights are reshaped by a Hebbian rule that reinforces correlation between new-view and past representations, combined with random binary masking of first-layer synapses (synaptic partition learning) and a weight-drift penalty to retain old knowledge.
Key results
- MVIL reaches 91.50% ACC on 100leaves and 95.04% on NGs, outperforming all listed baselines including the static multi-view methods TMC, LGCNFF and IHGCN in most datasets.
- On Flower17, MVIL reports 60.54% ACC, about 1.6 points above the second-best method (per the paper's text).
- Ablation on 100leaves shows the full model (C1+C2+C3) reaching 91.36% ACC versus 90.79% for C1+C2 alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add structured Hebbian plasticity (C3) to C1+C2 (ablation) | MVIL · 100leaves, Animals, Flower17, NGs, NoisyMNIST_15000, YaleB_Extended | ACC: 90.79 (0.69) -> 91.36 (0.68); 84.01 (0.11) -> 84.42 (0.10); 56.21 (0.42) -> 57.79 (0.37); 94.37 (0.42) -> 95.04 (0.64); | Helped |
| Add weights reconstructed loss (C2) to synaptic partition learning (C1) (ablation) | MVIL · 100leaves | ACC: 90.67 (0.17) -> 90.79 (0.69) | Helped |
| Add weights reconstructed loss (C2) to synaptic partition learning (C1) (ablation) | MVIL · Animals, Flower17, NGs, YaleB_Extended | ACC: 79.97 (0.06) -> 84.01 (0.11); 36.00 (0.21) -> 56.21 (0.42); 83.33 (0.65) -> 94.37 (0.42); 34.70 (0.58) -> 62.66 (0.16) | Helped |
| Add weights reconstructed loss (C2) to synaptic partition learning (C1) (ablation) | MVIL · NoisyMNIST_15000 | ACC: 95.05 (0.64) -> 91.91 (0.12) | Hurt |
| Add synaptic partition learning (C1) to weights reconstructed loss (C2) (ablation) | MVIL · 100leaves, Animals, Flower17, NGs, NoisyMNIST_15000, YaleB_Extended | ACC: C1+C2: 90.79 (0.69), 84.01 (0.11), 56.21 (0.42), 94.37 (0.42), 91.91 (0.12), 62.66 (0.16); C2: 86.79 (0.20), 80.00 (0.05 | Mixed |
| Use MVIL instead of the strongest reported baseline | MVIL · 100leaves, Animals, Flower17, YaleB_Extended | ACC: MVIL: 91.50 (0.47), 84.42 (0.10), 60.54 (0.72), 66.99 (0.28); strongest baseline: 91.03 (0.70), 83.13 (0.04), 58.94 (0.1 | Helped |
| Use MVIL instead of IHGCN | MVIL · NGs and NoisyMNIST_15000 | ACC: IHGCN: 96.89 (0.14), 93.64 (0.46); MVIL: 95.04 (0.64), 92.43 (0.64) | Hurt |
Surprising findings
- On NoisyMNIST_15000, single-component C1 alone (95.05%) outperforms the full model (92.43%) in ACC, and the paper's own knowledge-accumulation analysis notes that single-view learning sometimes beats MVIL.
- Static multi-view method TMC achieves the best recall on Animals (79.19%) and the best precision on YaleB_Extended, beating MVIL on those metrics.
Limitations
- Synaptic partition masking can limit performance, so single-view learning sometimes outperforms MVIL on some datasets.
- Evaluated only on node classification over graph-based multi-view datasets; classification beyond graphs is not tested.
- Reported gains are within reported standard deviations in several cases and rely on 3 repeated runs.
Experimental setup
Datasets: 100leaves Animals Flower17 NGs NoisyMNIST_15000 YaleB_Extended
Benchmarks: 100leaves Animals Flower17 NGs NoisyMNIST_15000 YaleB_Extended
Metrics: ACC Precision Recall Macro-F1
Baselines: GAT SI MAS LGCNFF IHGCN TMC MVCIL
Builds on: Hebbian learning graph convolutional network synaptic intelligence memory aware synapses
Closest papers by meaning
- Learning the Plasticity: Plasticity-Driven Learning Framework in Spiking Neural Networks (NeurIPS 2025 · similarity 0.86)
- Biologically Inspired Learning Model for Instructed Vision (NeurIPS 2024 · similarity 0.85)
- Flexible Context-Driven Sensory Processing in Dynamical Vision Models (NeurIPS 2024 · similarity 0.85)
- MoRE-Brain: Routed Mixture of Experts for Interpretable and Generalizable Cross-Subject fMRI Visual Decoding (NeurIPS 2025 · similarity 0.85)
- SynBrain: Enhancing Visual-to-fMRI Synthesis via Probabilistic Representation Learning (NeurIPS 2025 · similarity 0.84)
- Rethinking Hebbian Principle: Low-Dimensional Structural Projection for Unsupervised Learning (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.