| name | intrinsic-noise-consolidation-doob |
| description | Doob-Barrier-Conditioned Diffusion methodology that turns analog neuromorphic device noise into a continual-learning resource. Casts per-synapse consolidation as a Doob h-transform, creating a noise-amplified restoring force that consolidates memories — predicting an inverted-U relationship between noise level and sequential-task retention. |
| version | 1.0.0 |
| author | Hermes Agent |
| license | MIT |
| metadata | {"hermes":{"tags":["neuromorphic","continual-learning","doob-h-transform","device-noise","brainscales2","memory-consolidation","analog-hardware","stability-plasticity"],"category":"ai_collection","arxiv_id":"2607.06924","arxiv_url":"https://arxiv.org/abs/2607.06924","published":"2026-07-08","authors":["Gunner Levi Howe"],"categories":["cs.LG","cs.NE"],"trigger_words":["doob barrier","h-transform","intrinsic noise","consolidation","analog noise","neuromorphic hardware","brainscales","continual learning","stability-plasticity","inverted-u","device noise"]}} |
| created | 2026-07-12 |
| updated | 2026-07-12 |
Intrinsic-Noise Consolidation: A Doob-Barrier-Conditioned Diffusion Turns Analog Device Noise into a Continual-Learning Resource
arXiv: 2607.06924 | Published: 2026-07-08 | Author: Gunner Levi Howe
Core Thesis
On analog neuromorphic hardware, intrinsic device noise is normally treated as an accuracy tax. This paper asks: can it instead consolidate memories?
The key insight: cast per-synapse consolidation as a Doob h-transform — condition each weight's stochastic dynamics on never crossing a memory-critical barrier around its consolidated value.
The Math
The conditioned diffusion gains an extra drift term:
σ² · d/dw log h(w)
This is a restoring force amplified by the noise variance itself that diverges at the barrier.
Key Novelty Claims
- Doob barrier-conditioning as a synaptic rule — every h-transform use found in literature is for generative modeling, none for synaptic consolidation
- Falsifiable prediction: Increasing intrinsic noise non-monotonically improves sequential-task retention — an inverted-U that anchored-drift methods (OUA, MESU, EWC) cannot produce
Experimental Results
Simulation (Split-MNIST, 8 seeds)
- The rule lifts retention 10.9 points at an interior optimum (paired Wilcoxon p=0.004)
- Matched OU/EWC/MESU anchors are monotone (no inverted-U)
- Ablating the conditioning removes the effect
- The optimum tracks the barrier
- The inverted-U survives a second task stream and forward-pass noise
Hardware (BrainScaleS-2)
- Measured intrinsic noise on real silicon (additive, trial-to-trial independent, tunable via on-chip averaging)
- Barrier-conditioning retains prior task 15.6 points better than matched control at matched average accuracy
- This is a stability-plasticity shift, not a net-accuracy win
- Single seed; retention measured, energy modeled
Key Insight
Intrinsic analog noise becomes a consolidation dividend — a digital accelerator must spend energy to generate what analog hardware gets for free.
Practical Applications
1. Neuromorphic Continual Learning
- Use device noise as a feature, not a bug
