Canonical Result — 50 Exemplars/Class Replay
Average Accuracy
16.03%
Condition E · Viveka gain
+9.21pp vs baseline
Backward Transfer
−50.5%
BWT · forgetting rate
+11.5pp vs baseline
Viveka vs Smriti
+0.63pp
AA delta · E vs C
Consistent gain
Ortho Head Penalty
−8.84pp
AA delta · D vs C
Vairagya collapse
Bhaya Quiescence
0.000
Tasks 1–9 · all replay conditions
Replicated from P4
Buddhi Lock Epoch
Ep. 5
Stable across all conditions
Design determinism
Replay Budget Effect
20 vs 50 exemplars/class — confirms replay budget as primary CIL bottleneck
Within-Task Learning Strength
Final-epoch confidence per task (Run 2) — strong learning despite poor cross-task retention
Six-Condition Ablation Study
| Condition | Replay | Vairagya | Buddhi | Viveka Gain | Ortho Head | AA (%) | BWT (%) | ΔAA vs Baseline | Vairagya Sat. |
|---|---|---|---|---|---|---|---|---|---|
| A · SGD Baseline | ✗ | ✗ | ✗ | ✗ | ✗ | 6.82 | −62.01 | — | — |
| B · Replay Only | ✓ | ✗ | ✗ | ✗ | ✗ | 15.32 | −53.62 | +8.50 | ~0.59 |
| C · Maya-Smriti | ✓ | ✓ | ✓ | ✗ | ✗ | 15.40 | −50.52 | +8.58 | ~0.47 |
| D · Maya-Smriti + Ortho | ✓ | ✓ | ✓ | ✗ | ✓ | 6.56 | −60.38 | −0.26 | ~0.24 ⚠ |
| E · Viveka, learnable head ★ | ✓ | ✓ | ✓ | ✓ | ✗ | 16.03 | −50.50 | +9.21 | ~0.47 |
| F · Full Maya-Viveka (E+D) | ✓ | ✓ | ✓ | ✓ | ✓ | 6.66 | −60.24 | −0.16 | ~0.27 ⚠ |
★ = Canonical result
⚠ = Vairagya collapse — prototype geometry suppresses consolidation
Accuracy Matrices R[trained_up_to][task_id]
Condition E — Maya-Viveka (Viveka gain, learnable head) · AA=16.03% · BWT=−50.50%
Lower-triangular: diagonal = within-task accuracy, off-diagonal = cross-task retention. Orange = meaningful retention; grey = zero (unseen tasks).
Condition B — Replay Only · AA=15.32% · BWT=−53.62%
Baseline comparison: replay alone without affective mechanisms.
Condition A — SGD Baseline · AA=6.82% · BWT=−62.01%
No replay, no affective system. Pure catastrophic forgetting — diagonal only.
Affective Dynamics
Bhaya Quiescence — Pain Signal Suppression Under Replay
Mean Bhaya firing rate per task. Baseline escalates; all replay conditions quiesce to exactly 0.000 after Task 0 — emergent property replicated from P4.
Emergent finding: Bhaya fires only during Task 0 (rate=0.024) across every replay condition, then drops to exactly 0.000 and stays there. Baseline continues escalating to 0.12+ as forgetting compounds. Consistent with P4 — a calm, replay-supported network does not experience nociceptive pain.
Buddhi S-Curve — Consolidation Gate Recovery
Mean Buddhi per epoch across inter-task training (Tasks 1–9). Collapses at task boundary, recovers with experience. Identical across all conditions — architecturally deterministic.
Design confirmation: Buddhi S-curve (0.10→0.30→0.50→0.70→0.90 at epochs 0–4, locks at 1.0 from epoch 5) is identical across all 6 ablation conditions. Buddhi is architecture-stable regardless of head type or Viveka presence.
Vairagya Saturation — Ortho Head Collapse
Mean Vairagya (final epoch) per task across key conditions. Ortho head suppresses saturation from ~0.47 to ~0.24 — the mechanistic cause of AA collapse in conditions D and F.
V-fc1 Protection Trajectory
Fraction of fc1 synapses under Vairagya protection (final epoch per task). Viveka extends protection further and decays more slowly than Maya-Smriti alone.
Viveka Gain Trajectory — Cross-Task Synaptic Consistency (Run 2, 50/class)
Mean Viveka score (final epoch) per task. Rises from 0.0 as the network accumulates cross-task experience, stabilises at 0.50–0.58 range. Viveka is a learned signal, not a fixed prior.
Bhaya Quiescence Heatmap
Mean Bhaya Firing Rate per Task — All Replay Conditions
Green = quiescent (0.000). Orange/red = active firing. Baseline escalates; replay conditions show perfect quiescence from Task 1 onwards.
Key Findings
Finding 01
Viveka Gain Provides Consistent Improvement
Viveka-gated selective synaptic protection adds a consistent +0.63pp AA over Maya-Smriti on the harder Split-CIFAR-100 benchmark. The gain is modest but mechanistically clean — cross-task consistency scores discriminate stable from unstable synapses, extending V-fc1 protection by ~2–4pp per task.
E vs C: +0.63pp AA | V-fc1 peaks at 56% vs 51%
Finding 02
Orthogonal Head Causes Vairagya Collapse
Adding an orthogonal prototype head suppresses Vairagya saturation from ~0.47 to ~0.24 — a 49% reduction. The prototype geometry actively fights the affective consolidation signal: as orthogonal constraints rigidify the output space, Vairagya cannot protect synapses freely. AA collapses from 15.40% back to 6.56% — near baseline.
D vs C: −8.84pp AA | Vairagya: 0.47→0.24
Finding 03
Bhaya Quiescence Replicates on Harder Benchmark
First confirmed in P4 on Split-CIFAR-10, Bhaya quiescence now replicates on Split-CIFAR-100 — a 10× harder class space. Every replay condition shows Bhaya firing only during Task 0 (rate=0.024), then exactly 0.000 for Tasks 1–9. Baseline continues escalating. This is a robust emergent property of replay-stabilised affective SNNs.
Bhaya: 0.024 (T0) → 0.000 (T1–T9) across all replay conditions
Finding 04
Buddhi S-Curve is Architecturally Deterministic
The Buddhi consolidation gate follows an identical S-curve across all 6 ablation conditions: 0.10, 0.30, 0.50, 0.70, 0.90 at epochs 0–4, locking at 1.0 from epoch 5 onwards. Head type and Viveka presence have zero effect on Buddhi dynamics. This confirms Buddhi is a stable architectural prior, not an adaptive signal.
Buddhi curve: identical across A, B, C, D, E, F
Finding 05
Replay Budget Remains Primary Bottleneck
Moving from 20 to 50 exemplars/class yields +4.10pp AA and +4.96pp BWT — larger than any architectural addition in the ablation. Viveka and Vairagya provide genuine signal above this floor, but the replay budget sets the ceiling. This motivates P6: a retrograde consolidation mechanism that modulates what was consolidated rather than requiring more exemplars.
20→50/class: +4.10pp AA | Viveka: +0.63pp additional
Finding 06 · P6 Motivation
Orthogonal Collapse Reveals Missing Retrograde Signal
The orthogonal head failure is not random noise — it reveals a specific architectural gap. The network consolidates synapses forward during learning (Vairagya protection), but has no mechanism to revise that decision after a task boundary when prototype geometry shifts. This is exactly the endocannabinoid retrograde signalling problem. P6 addresses it directly.
P6: retrograde consolidation correction via eCB-inspired signal
Vedantic Architecture — Antahkarana Mapping
Each affective dimension corresponds to a faculty of the Antahkarana (inner instrument) in Advaita Vedanta. The series maps the complete cognitive architecture onto independently falsifiable SNN components.
| Dimension | Sanskrit | Computational Role | τ / Mechanism | Status | Paper |
|---|---|---|---|---|---|
| Bhaya | भय | Fear · nociceptive pain trigger · gradient signal | τ = 3 | Active P1–P5 | P1 |
| Vairagya | वैराग्य | Wisdom · heterosynaptic decay gating · selective consolidation | τ = 20 | Active P1–P5 | P1 |
| Shraddha | श्रद्धा | Trust · confidence integrator | τ = 10 | Active P1–P5 | P1 |
| Spanda | स्पन्द | Aliveness · spike rate monitor | τ = 5 | Active P1–P5 | P1 |
| Buddhi | बुद्धि | Intellect · S-curve consolidation gate · Viparita Buddhi at task boundaries | τ = 200 | Active P4–P5 | P4 |
| Viveka ★ | विवेक | Discernment · cross-task synaptic consistency · Vairagya gain modulation | Learned score | New · P5 | P5 |
| Ahamkara | अहंकार | Ego · task-attachment as forgetting cause · fc_out lock-out | Failure mode | Named P4 · Dissolved P5 | P4→P5 |
| Samskara | संस्कार | Impression traces · cross-task synaptic memory | TBD | Planned P6 | P6 |
| Chitta | चित्त | Implicit synaptic memory · retrograde eCB signal | TBD | Planned P6 | P6 |
| Prana | प्राण | Metabolic plasticity budget · embodied energy constraint | TBD | Planned P9 | P9 |
Series thesis: The Antahkarana (inner cognitive instrument) in Advaita Vedanta maps precisely onto a neuromorphic CIL architecture. Each paper introduces exactly one new Vedantic dimension, enabling clean ablation and isolated contribution claims. The complete system — Manas (P1), Buddhi (P4), Viveka (P5), Chitta (P6), Prana (P9) — constitutes the Antahkarana in SNN form.
Maya Research Series