Maya-Viveka: Viveka-Gated Synaptic Discrimination for Class-Incremental Learning in Affective Spiking Neural Networks

Maya-Viveka extends the Maya affective SNN architecture to Class-Incremental Learning on Split-CIFAR-100 (10 tasks, 10 classes each, 100 total classes) through the introduction of Viveka (विवेक) — discernment — as a sixth affective dimension. Viveka is a cross-task synaptic consistency tracker, inspired by the GANE (Gain-Amplified Neural Encoding) norepinephrine model, that modulates Vairagya protection strength based on per-synapse representational stability across task boundaries. Synapses encoding features consistently activated across multiple tasks receive elevated Vairagya protection; synapses encoding task-specific transient activations receive standard protection. A six-condition ablation study establishes that Viveka-gated selective protection (Condition E) achieves AA=16. 03%, BWT=−50. 50%, outperforming the SGD baseline (AA=6. 82%) by +9. 21 pp and Maya-Smriti alone (AA=15. 40%) by +0. 63 pp. A critical failure mode is identified: orthogonal prototype enforcement at insufficient replay budget actively suppresses Vairagya saturation from ~0. 47 to ~0. 24, causing representational collapse to baseline-equivalent performance (AA=6. 56%). This orthogonal collapse finding reveals a structural gap — the absence of a retrograde consolidation correction signal — that directly motivates Paper 6's endocannabinoid-inspired mechanism. Affective quiescence — the suppression of Bhaya (fear) to exactly 0. 000 across Tasks 1–9 in all replay conditions — is confirmed as a replicable emergent property on a benchmark ten times harder than Paper 4. Buddhi's S-curve consolidation gate is architecturally stable and identical across all six ablation conditions, confirming design determinism. Paper 5 in the Maya Research Series. GitHub: https: //github. com/venky2099/Maya-Viveka. Interactive results dashboard: https: //venky2099. github. io/Maya-Viveka/mayaᵥivekadashboard. html