| name | intrinsic-noise-consolidation-continual-learning |
| description | Doob-barrier-conditioned diffusion methodology for turning analog neuromorphic hardware noise into a continual-learning resource — per-synapse consolidation via Doob h-transform creates noise-amplified restoring force that yields inverted-U noise-retention relationship. Validated on BrainScaleS-2 neuromorphic silicon with hardware-in-the-loop training. |
| category | neuroscience |
Intrinsic-Noise Consolidation for Continual Learning
Overview
This skill implements the Doob-barrier-conditioned diffusion methodology from arXiv:2607.06924v1, which reframes analog neuromorphic hardware noise from an accuracy tax into a consolidation dividend. The core insight: conditioning synaptic weight diffusion on never crossing a memory-critical barrier creates a noise-amplified restoring force that non-monotonically improves sequential-task retention.
Key paper: Intrinsic-Noise Consolidation: A Doob-Barrier-Conditioned Diffusion Turns Analog Device Noise into a Continual-Learning Resource (Howe — 2026-07-08)
Core Methodology
Doob h-Transform for Synaptic Consolidation
After learning a task, a synaptic weight w should stay near consolidated value μ. Model as diffusion:
Unconditioned (OU process): dw = -s(w-μ)dt + σdW
- Stationary spread σ²/2s grows with noise → more noise is strictly worse
Barrier-conditioned (Doob h-transform): Condition diffusion on never crossing barrier at μ±b
- Conditioned process gains extra drift: σ²∂_w log h(w)
- Where h(w) is the survival probability (never hitting barrier)
- This drift: (i) points into safe region, (ii) diverges at barrier, (iii) scales with σ²
The Inverted-U Prediction
The σ² steering term (stronger at moderate noise) competes with raw σ diffusion (overwhelming at high noise):
Retention ▲
│ ● optimal σ*
│ / \
│ / \
│ / \
│ / \
│/ \
└──────────────► Noise (σ)
This predicts a non-monotonic inverted-U relationship between noise and retention — a falsifiable prediction that anchored-drift methods (OU, EWC, MESU) cannot produce.
Key Intellectual Claim Split
The method has three components; intellectual honesty requires separating them:
- Surrendered (known): The anchored drift
-s(w-μ) — identical to small-noise limit of OU Adaptation, variance-scaled anchor of MESU, Fisher penalty of EWC
- Novel (a): Doob barrier-conditioning as a synaptic rule — every h-transform use found was generative modeling/Schrödinger bridges, none synaptic
- Novel (b): Falsifiable hardware curve prediction — increasing intrinsic noise non-monotonically improves retention
Consolidation Rule
dw = [-s(w-μ) + σ²∂_w log h(w)]dt + σdW
↑ ↑
anchored drift Doob barrier-conditioning
(known) (novel, noise-amplified)
Experimental Validation
E0 — Pre-registered Falsifier
- Single-head Split-MNIST, 8 seeds
- Barrier-conditioned rule lifts retention by 10.9 percentage points at σ* = 0.02
- Paired Wilcoxon p = 0.004 vs. zero noise and vs. high noise
- Matched OU, EWC, MESU anchors are monotone-decreasing in noise
E2 — BrainScaleS-2 Noise Model
- The inverted-U survives device-faithful BSS-2 noise model:
- Colored noise
- Multiplicative noise
- Fixed-pattern noise
- 6-bit quantization
E5 — On-Silicon Noise Measurement
- Real BrainScaleS-2 silicon (chip hxcube7fpga3chip61_1)
- Additive, trial-to-trial-independent noise
- Coefficient of variation up to 12.0%
num_sends knob averages as ≈ 1/√N — the benign noise class the mechanism needs
E7 — On-Silicon Hardware-in-the-Loop
- Real BrainScaleS-2 silicon with chip in training loop
- Intrinsic noise steered by barrier-conditioning retains prior task 15.6 points better than matched unconditioned control
- Single seed, one operating point; retention measured, energy modelled
Implementation Patterns
Barrier-Conditioned Synaptic Update
def doob_barrier_drift(w, mu, barrier, sigma):
"""Compute σ²∂_w log h(w) — the Doob barrier-conditioning drift."""
pass
def consolidated_update(w, mu, sigma, barrier, s):
"""Full barrier-conditioned synaptic update."""
anchored = -s * (w - mu)
doob = doob_barrier_drift(w, mu, barrier, sigma)
noise = sigma * random_normal()
return anchored + doob + noise
Finding the Optimal Noise Level
for sigma in noise_range:
retention = train_with_noise(sigma)
if retention > best_retention:
best_sigma = sigma
best_retention = retention
BrainScaleS-2 Integration
Key Insights
- Noise as resource, not tax: On analog neuromorphic hardware, intrinsic noise can be steered to consolidate memories rather than degrade them
- Hardware advantage: Von-Neumann accelerators must spend energy to generate noise that neuromorphic hardware has for free
- Falsifiable prediction: The inverted-U noise-retention curve is a load-bearing prediction — if it doesn't appear, the mechanism doesn't work
- Device-specific tuning: The optimal σ* depends on the device's noise profile and the barrier width
Pitfalls
- Barrier width must be task-appropriate: Too narrow → weights can't learn new tasks; too wide → no consolidation benefit
- Noise must be additive and trial-independent: Multiplicative or correlated noise may not produce the inverted-U
- The anchored drift is NOT novel: Don't claim novelty for
-s(w-μ) — it's known from OU/MESU/EWC
- Single-seed hardware results: The on-silicon demo (E7) is single-seed — needs replication
- Energy modelling, not measurement: The hardware energy savings are modelled, not directly measured
Activation Keywords
Doob h-transform, barrier-conditioned diffusion, intrinsic noise consolidation, continual learning, catastrophic forgetting, BrainScaleS-2, neuromorphic hardware, analog noise, synaptic consolidation, inverted-U noise retention, Ornstein-Uhlenbeck, MESU, EWC, rehearsal-free learning