Doob Barrier Noise Consolidation

---
name: doob-barrier-noise-consolidation
description: >
  Doob-barrier-conditioned diffusion methodology that turns analog device
  noise into a continual-learning resource for neuromorphic hardware. Uses
  Doob h-transform to condition synaptic weight dynamics on never crossing
  memory-critical barriers, creating a noise-amplified restoring force.
  Based on arXiv:2607.06924v1 (Howe, 2026).
---

# Intrinsic-Noise Consolidation via Doob-Barrier-Conditioned Diffusion

## Source

arXiv: 2607.06924v1 — "Intrinsic-Noise Consolidation: A Doob-Barrier-Conditioned Diffusion Turns Analog Device Noise into a Continual-Learning Resource"
Author: Gunner Levi Howe
Date: 2026-07-08
Categories: cs.LG, cs.NE

## Core Concept

On analog neuromorphic hardware (e.g., BrainScaleS-2), intrinsic device noise is normally treated as an accuracy tax. This paper reframes it as a **consolidation dividend**: by conditioning each weight's stochastic dynamics on never crossing a memory-critical barrier around its consolidated value, the noise itself generates a restoring force toward the memory.

## Mathematical Framework

### Doob h-Transform for Synaptic Consolidation
- Cast per-synapse consolidation as a Doob h-transform
- Condition each weight's diffusion on the event of never crossing a barrier around its consolidated value
- The conditioned diffusion acquires an extra drift: σ² · ∂w log h
- This restoring force is **amplified by the noise variance itself** and diverges at the barrier

### Key Equation
The barrier-conditioned synaptic weight dynamics:
```
dw = -s(w - μ)dt + σ² · ∂w log h · dt + σ dW
```
Where:
- `-s(w - μ)` = anchored-consolidation drift (re-derivation of OU Adaptation, MESU, EWC)
- `σ² · ∂w log h` = **novel Doob barrier-conditioning term** (main contribution)
- `σ dW` = intrinsic device noise

### Falsifiable Prediction
Increasing intrinsic noise **non-monotonically** improves sequential-task retention (inverted-U curve) — a pattern that anchored-drift methods (OU, EWC, MESU) cannot produce.

## Experimental Results

### Pre-Registered Gate (E0)
- Single-head Split-MNIST (8 seeds)
- Barrier-conditioned rule lifts retention by **10.9 percentage points** at interior optimum σ* = 0.02
- Paired Wilcoxon p = 0.004 vs. zero noise and vs. high noise
- Matched OU, EWC, and MESU anchors are monotone-decreasing in noise

### Mechanism Isolation (E1)
- Ablating the conditioning removes the effect
- Optimum tracks the barrier position
- Inverted-U survives across second task stream

### Hardware-Faithful Validation
- Survives BrainScaleS-2 noise model (colored, multiplicative, fixed-pattern, 6-bit)
- Survives hardware-faithful realization where noise enters forward pass (analog MAC) rather than weights
- Retention optimum tunable to device's few-percent intrinsic noise

### On-Silicon Demonstration (E7)
- Real BrainScaleS-2 silicon measurement: additive, trial-to-trial-independent noise
- Coefficient of variation up to 12.0%
- Chip's num_sends knob averages as ≈ 1/√N
- On-silicon run: retains prior task **15.6 points better** than matched unconditioned control at matched average accuracy

## Comparison to Existing Methods

| Method | Noise Dependency | Retention Mechanism |
|--------|-----------------|---------------------|
| EWC | Monotone decreasing | Fisher penalty anchor |
| OU Adaptation | Monotone decreasing | Ornstein-Uhlenbeck anchor |
| MESU | Monotone decreasing | Variance-scaled anchor |
| **Doob Barrier** | **Inverted-U (optimal σ*)** | **Noise-amplified restoring force** |
| Replay (with data storage) | N/A | Direct rehearsal |

At its optimum, the Doob barrier rule is the **strongest rehearsal-free consolidation method** tested — matching MESU and significantly beating OU and EWC.

## Key Insights

1. **Noise as resource, not tax**: On von-Neumann accelerators, stochasticity must be added deliberately and paid for. On analog neuromorphic substrates, noise is free — this mechanism harnesses it.

2. **Stability-plasticity shift**: The mechanism trades net accuracy for stability, not a pure accuracy win.

3. **Device-faithful**: Works with real hardware noise characteristics (colored, multiplicative, fixed-pattern, quantized).

4. **Pre-registered falsifier**: The inverted-U prediction was pre-registered as a go/no-go gate and passed.

## Applications

### Continual Learning on Neuromorphic Hardware
- BrainScaleS-2, Loihi, TrueNorth, other analog/digital neuromorphic chips
- Sequential task learning without rehearsal data storage
- Energy-efficient memory consolidation

### Analog Hardware Design
- Guides noise tolerance specifications for neuromorphic chips
- Informs optimal operating points for analog device noise levels
- Reframes device noise from liability to computational resource

### Catastrophic Forgetting Mitigation
- Alternative to EWC, replay, and other consolidation methods
- Particularly effective when noise amplitude matches barrier width
- Complements existing anchored-drift approaches

## Trigger Words

doob h-transform, noise consolidation, continual learning, catastrophic forgetting, BrainScaleS-2, neuromorphic noise, analog hardware, synaptic consolidation, barrier-conditioned diffusion, stability-plasticity, EWC alternative, rehearsal-free learning