Controlled
comparison
Two confirmed tasks. Every architecture is trained and evaluated per seed under identical data conditions; the plots show mean accuracy with ±1 SD across seeds.
Binary Pattern DiscriminationStatic pattern
A static, globally observable binary classification task. The model receives a single pattern and must assign it to one of two classes. There is no temporal structure and no delay between stimulus and decision.
Range marks show ±1 SD across seeds. All metrics (95% CI, parameter counts, seed counts, training times) are sourced from real NeuroWeave experiment artifacts. Click any row to expand.
Simple static global classification does not reveal a unique MaleCNS advantage — dense, random-graph and conventional baselines reach the same ceiling.
Multi-seed evaluation (5 seeds, 40 epochs). Each architecture is trained and evaluated independently per seed; the reported value is the mean accuracy with the standard deviation across seeds. 95% CI computed via t-distribution.
Temporal Sequence MemoryTemporal memory
A delayed-recall task with temporal structure: information presented early in a sequence must be retained across intervening steps before a decision is required. This probes whether a computational architecture supports useful state retention rather than instantaneous mapping.
Range marks show ±1 SD across seeds. All metrics (95% CI, parameter counts, seed counts, training times) are sourced from real NeuroWeave experiment artifacts. Click any row to expand.
The MaleCNS-derived architecture with SA-010 temporal dynamics shows strong performance on this specific delayed-recall task, while the conventional baselines evaluated here do not.
Multi-seed evaluation (5 seeds, 30 epochs) of the final confirmed configuration. Temporal dynamics are parameterised by an activation decay and a leak term applied to the connectome-constrained layer. 95% CI computed via t-distribution.
The SA-010 decay/leak setting (decay 0.05 / leak 0.02) was selected through exploratory sensitivity analysis on this task. It should be read as a task-specific configuration finding, not as a universal biological conclusion. Confidence intervals are wide relative to the differences between conditions.