Autonomous Continual Learning and Knowledge Transfer via Cortex-Inspired Neural Assemblies
Abstract
Unlike the brain, artificial neural networks (ANNs) suffer from "catastrophic forgetting": complete overwriting of solutions when re-training on new functions ("concept shift") or on a disjoint domain of the current function ("domain shift"). ANNs require each batch of data to be roughly representative of the complete distribution, a typically expensive and impractical requirement. In this paper, we introduce Functional Task Networks (FTNs), a method of continual learning and knowledge transfer modeled after several properties of the mammalian neocortex. FTN is formulated as a distributed, stochastic mixture of experts in which a sparse, binary routing mask is computed at inference time using a small support sample of input-output pairs. Per-batch reconfiguration of the router during training, together with mask-overlap-filtered replay, shapes a basis of overlapping, partially trained subnetworks. Unrelated tasks route through mostly separate assemblies, related tasks overlap on shared components. On block-sequentially trained continual learning and knowledge transfer tasks FTN retains solutions nearly as well as disjoint-mask isolation and attains the lowest pre-update head-start loss among matched-architecture comparators. The learned mask-overlap geometry tracks the known concept-sharing structure where available. We show experimentally that under correlated concept shift, support-set mask reconfiguration is necessary even for retention. Under domain shift, the distributed bank with replay is sufficient to handle retention without the configurer. Forward transfer, or head-start, and backward transfer, are also contributed by support-set mask configuration.
Figures & tables
| Method | ACC | FM | BWT | HS |
|---|---|---|---|---|
| NoMask (naive seq.) | ||||
| NoMask + ER (random replay) | ||||
| FixedMask (oracle ID at eval) | ||||
| FixedMask + retrieval router | ||||
| FTN (ours) | ||||
| FTN, random masks (ablation) |
| Model | Routing | Replay | Recall | Forward | FT change | Train | Current | Overlap (%) |
|---|---|---|---|---|---|---|---|---|
| MLP | Fixed | Filtered | 0.922 | 0.785 | 0.0118 | 0.950 | 0.941 | 15.78 |
| MLP | Fixed | None | 0.805 | 0.735 | 0.0055 | 0.954 | 0.946 | 14.76 |
| MLP | Changepoint-Driven | Filtered | 0.912 | 0.771 | 0.0107 | 0.952 | 0.923 | 13.45 |
| MLP | Changepoint-Driven | None | 0.886 | 0.756 | 0.0111 | 0.947 | 0.929 | 12.40 |
| LoRA | Fixed | Filtered | 0.906 | 0.780 | 0.0093 | 0.908 | 0.916 | 16.93 |
| LoRA | Fixed | None | 0.656 | 0.663 | 0.0012 | 0.825 | 0.853 | 40.38 |
| Shared-shuffle | Permuted | Rotated | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Method | ACC | FM | HS | ACC | FM | HS | ACC | FM | HS |
| NoMask | 0.444 | 0.644 | 10.397 | 0.928 | 0.053 | 5.207 | 0.523 | 0.509 | 0.360 |
| NoMask+ER | 0.442 | 0.630 | 8.524 | 0.960 | 0.018 | 5.270 | 0.912 | 0.075 | 0.667 |
| SupSup | 0.886 | 0.000 | 3.535 | 0.894 | 0.000 | 2.683 | 0.894 | 0.000 | 0.861 |
| Progressive Nets | 0.979 | 0.000 | 8.330 | 0.977 | 0.000 | 3.096 | 0.977 | 0.000 | 0.381 |
| FixedMask | 0.979 | 0.000 | 2.309 | 0.979 | 0.000 | 2.321 | 0.979 | 0.000 | 2.322 |
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
| Method | ACC | FM | BWT | Head-start | |
|---|---|---|---|---|---|
| NoMask | 0.610 0.019 | 0.392 0.016 | -0.392 0.016 | 2.830 0.293 | 0.000 0.000 |
| FixedMask | 0.945 0.020 | 0.000 0.000 | +0.000 0.000 | 0.702 0.003 | 0.000 0.000 |
| FTN (AnnealSTE, ) | 0.886 0.042 | 0.055 0.015 | -0.053 0.017 | 0.423 0.077 | 0.454 0.213 |
| Method | ACC | FM | BWT | Head-start | |
|---|---|---|---|---|---|
| NoMask | 0.433 0.056 | 0.656 0.067 | -0.656 0.067 | 10.840 0.881 | 0.000 0.000 |
| FixedMask | 0.978 0.001 | 0.000 0.000 | +0.000 0.000 | 2.312 0.016 | 0.000 0.000 |
| FTN (AnnealSTE, ) | 0.962 0.004 | 0.021 0.005 | -0.021 0.005 | 0.573 0.039 | 0.398 0.127 |
| Method | Budget | ACC | FM | BWT | Head-start | |
|---|---|---|---|---|---|---|
| Sigmoid | 10 | 0.900 | 0.057 | -0.057 | 0.484 | 0.491 |
| Sigmoid | 20 | 0.897 | 0.057 | -0.057 | 0.432 | 0.581 |
| Sigmoid | 30 | 0.894 | 0.061 | -0.061 | 0.444 | 0.477 |
| Sigmoid | 40 | 0.920 | 0.040 | -0.040 | 0.422 | 0.226 |
| Sigmoid | 50 | 0.914 | 0.046 | -0.046 | 0.494 | 0.221 |
| Sigmoid | 60 | 0.912 | 0.050 | -0.050 | 0.462 | 0.262 |
| Method | ACC | FM | BWT | HS |
|---|---|---|---|---|
| NoMask (naive seq.) | ||||
| NoMask + ER (random replay) | ||||
| FixedMask (oracle ID at eval) | ||||
| FixedMask + retrieval router | ||||
| FTN, random per-batch masks (ablation) | ||||
| FTN (ours) |
| Method | ACC | FM | BWT | HS |
|---|---|---|---|---|
| NoMask (naive seq.) | ||||
| NoMask + ER (random replay) | ||||
| FixedMask (oracle ID at eval) | ||||
| FixedMask + retrieval router | ||||
| FTN, random per-batch masks (ablation) | ||||
| FTN (ours) |
| Method | ACC | FM | BWT | HS |
|---|---|---|---|---|
| NoMask (naive seq.) | ||||
| NoMask + ER (random replay) | ||||
| FixedMask (oracle ID at eval) | ||||
| FixedMask + retrieval router | ||||
| FTN, random per-batch masks (ablation) | ||||
| FTN (ours) |
| Condition | mean off-diag | max off-diag | chance ( ) | ||
|---|---|---|---|---|---|
| Task structure | |||||
| Correlated | 8 | 512 | |||
| Orthogonal | 8 | 512 | |||
| Reconfig freq | |||||
| per-batch | 8 | 512 | |||
| per-task | 8 | 512 | |||
Continual Learning
Continual Learning studies how models acquire knowledge from changing data streams while retaining or selectively revising what they learned before.
See all 323 papers