What question did this study set out to answer?

The aim is to introduce a novel gradient mechanism in spiking neural networks to improve continual learning efficiency.

April 1, 2026Open Access

Maya-Chitta: Endocannabinoid-Inspired Retrograde Gradient Gating for Class-Incremental Learning in Affective Spiking Neural Networks

Key Points

The aim is to introduce a novel gradient mechanism in spiking neural networks to improve continual learning efficiency.
Developed a retrograde gradient mechanism based on endocannabinoid signaling.
Implemented Chitta to accumulate historical task impressions during learning.
Applied gradient suppression to manage over-attachment to outdated knowledge.
Achieved a total Chitta gain of +0.16 percentage points.
Demonstrated +0.13 pp improvement in Moha boundary release.
Confirmed prior findings regarding structural calibration and dynamic behavior across tasks.

Abstract

Maya-Chitta introduces the first retrograde gradient mechanism in the Maya Research Series, a systematic mapping of Advaita Vedantic Antahkarana constructs onto neuromorphic spiking neural network (SNN) architecture for continual learning. The new dimension: Chitta (चित्त) In Advaita Vedanta, Chitta is the subconscious impression store — the substrate in which Samskaras (latent traces of past experience) reside and quietly shape present perception without being consciously retrieved. Computationally, Chitta is implemented as a per-synapse Samskara trace system that accumulates cross-task impression history during forward learning and applies retrograde gradient suppression to prevent Moha (pathological over-attachment to stale consolidations). The biological grounding is endocannabinoid (eCB) retrograde signalling (Wilson & Nicoll, 2001): post-synaptic neurons releasing retrograde messengers that travel back to pre-synaptic terminals to modulate release probability in proportion to usage history. Key results (Split-CIFAR-100 CIL, 10 tasks, seed=42): Gradient gate alone: +0. 03 pp AA over Maya-Viveka (Condition D vs C) Moha boundary release: +0. 13 pp AA (Condition E vs D) — dominant contribution Total Chitta gain: +0. 16 pp (AA=14. 42%, BWT=−53. 12%, Condition E canonical) NoGate = Full (E ≡ F): gate and Moha release are orthogonal at gate strength 0. 30 — a structural calibration datum for P7 Samskara traces active from Task 0, Epoch 1 (S=0. 011) — proactive, not reactive Bhaya Quiescence Law formally named: confirmed for the fifth consecutive paper across TIL CIFAR-10, CIL CIFAR-10, and three CIL CIFAR-100 experiments Buddhi S-curve determinism confirmed as series constant from P4 through P6 Series context: This is Paper 6 of the Maya Research Series. Previous papers: P1 Nociceptive Metaplasticity (DOI: 10. 5281/zenodo. 19151563), P2 Maya-OS (DOI: 10. 5281/zenodo. 19160123), P3 Maya-CL AA=62. 38% TIL (DOI: 10. 5281/zenodo. 19201769), P4 Maya-Smriti AA=31. 84% CIL (DOI: 10. 5281/zenodo. 19228975), P5 Maya-Viveka AA=16. 03% CIL (DOI: 10. 5281/zenodo. 19279002). Resources: GitHub: https: //github. com/venky2099/Maya-Chitta Interactive dashboard: https: //venky2099. github. io/Maya-Chitta/mayachittadashboardᵥ2. html ORCID: 0000-0002-3315-7907

Read Full Paperexternally

AI에게 질문

Bookmark

View Full Paper