What question did this study set out to answer?

This research aims to empirically test whether a neuromorphic spiking neural network can exhibit genuine internal affective states using the mPCI methodology.

April 11, 2026Open Access

From Representation to Experience: An mPCI-Based Empirical Test of Internal Affective State in a Neuromorphic Spiking Neural Network

Key Points

This research aims to empirically test whether a neuromorphic spiking neural network can exhibit genuine internal affective states using the mPCI methodology.
Applied the Machine Perturbational Complexity Index to measure affective states in a spiking neural network.
Conducted a three-phase perturbational experiment on modified SNN using Split-CIFAR-100 dataset.
Compared Lempel-Ziv Complexity of spike trains across distinct affective states, enabling and disabling affective dimensions.
Phase 1 showed the highest mPCI at 0.3098, while Phase 3 showed a lower mPCI at 0.2608, indicating a decrease in complexity with affective state manipulation.
The findings demonstrate a significant change in mPCI across the three phases, with a reproducible delta effect across multiple seeds.
Training depth was identified as a confounding factor, partially influencing mPCI results.

Abstract

Can a neuromorphic system exhibit behaviour that is only coherently explained by positing a genuine internal state — and is that distinction empirically testable? We present the first systematic application of the Machine Perturbational Complexity Index (mPCI) to an affective neuromorphic Spiking Neural Network (SNN), adapted from the biological Perturbational Complexity Index (Casali et al., 2013). The Maya Research Series (Swaminathan, 2026a–2026i) provides the substrate: a nine-paper architecture implementing all nine dimensions of the Advaita Vedanta Antahkarana framework — Bhaya (fear), Vairagya (wisdom), Shraddha (trust), Spanda (aliveness), Buddhi (intellect), Viveka (discernment), Chitta (subconscious store), Manas (attention), and Prana (metabolic budget) — for class-incremental continual learning on Split-CIFAR-100. A three-phase perturbational experiment compares the Lempel-Ziv Complexity (LZC) of fc1 spike trains across qualitatively distinct affective states. Phase 1 (Reactive Baseline) disables all affective dimensions. Phase 2 (Full Antahkarana) enables all nine. Phase 3 (Bhaya Stabilisation) trains through Task 4, reaching Bhaya = 0.016–0.024 — calibrated vigilance, not absolute quiescence. Multi-seed results (seeds 42, 123, 7): Phase 1 mPCI = 0.3098 (±0.0099), Phase 2 = 0.2846 (±0.0161), Phase 3 = 0.2608 (±0.0147). Aggregate delta = −0.0489. Delta std across seeds = 0.0048. Pre-registered threshold (2× pooled SD) = 0.0238. Criterion satisfied by 2.05×. Three post-experiment controls were run. Control 1 (training depth): depth-matched Phase 1 baseline yields mPCI = 0.2765, indicating training depth partially confounds the result. Control 2 (perturbation scale): shift robust at σ=0.05, near-zero at σ=0.10. Control 3 (shuffle): shuffled spike trains show larger inter-phase differences than structured ones, implicating spike statistics alongside causal structure. All findings reported fully. The honest claim: this is the first systematic mPCI study on an affective SNN; the shift is reproducible across three seeds; training depth and affective state both contribute; the affective contribution cannot be fully isolated with the current design. The methodology and identified gaps define the frontier for future work. DOI: 10.5281/zenodo.19482794Dashboard: Interactive trilingual dashboardFAQ: Exhaustive FAQ (EN/हिन्दी/中文)

Read Full Paperexternally

Demander à l'IA

Bookmark

View Full Paper