Noise Assisted Rare Event Simulation

---
name: noise-assisted-rare-event-simulation
description: "Noise-assisted internal simulation methodology. Deterministic attractor replay mis-represents rare events (greedy-recurrent over-representation OR low-prior-veto under-representation); moderate Ornstein-Uhlenbeck noise on unit support restores marginal AND conditional occurrence fidelity, with an inverted-U (stochastic resonance) signature and parameter-tolerance broadening."
license: MIT
metadata:
  arxiv_id: "2609.18033"
  published: "2026-09-16"
  authors: "Heng Zhang, Pawel Herman, Zenas C. Chao"
  affiliation: "WPI-IRCN UTokyo, KTH Royal Institute of Technology"
  tags: [neural-noise, internal-simulation, rare-events, bcpnn, stochastic-resonance, attractor-network, replay-fidelity, bayesian-confidence-propagation, predictive-processing, parkinsons-disease]
---

# Noise-Assisted Internal Simulation of Rare Events

**Methodology from arXiv:2609.18033** — how moderate neural noise enables a recurrent attractor network to faithfully express rare-event statistics learned from limited experience, without any retraining or parameter tuning at replay time.

## When to Use

- Modeling how brains estimate rare-event probabilities from finite samples (sampling error compensation)
- Designing generative replay / internal-simulation layers in predictive-processing architectures
- Diagnosing why a trained recurrent model collapses onto a few dominant trajectories and omits low-support alternatives
- Building noise-injection schedules for attractor networks that must reproduce a target distribution
- Studying altered neural variability (aging, Parkinson's, cholinergic blockade) as a drift off a corrective noise band

## Core Problem: Two Distinct Failure Stages

An internal model can fail at two separable stages:

1. **Learning stage**: finite-sample estimates of rare-event occurrence may under/overestimate ground truth (standard sampling problem).
2. **Expression stage**: even when evidence for rare events IS retained in learned parameters, deterministic autonomous dynamics may fail to express it.

This skill targets the **second, circuit-level stage**: how autonomous dynamics sample the learned structure. Key insight — replay fidelity depends not only on *what* is learned but on *how* the learned structure is sampled during internal simulation.

## Architecture: BCPNN Attractor Network

Bayesian Confidence Propagation Neural Network — recurrent attractor network with local Hebbian-Bayesian learning:

- **Units = minicolumns** grouped into one hypercolumn (neocortex-inspired); each unit represents one event; activity encodes confidence (probability) that event is active.
- **Support equation** (Ornstein-Uhlenbeck process on each unit):

```
ds_j = (1/τ_m) * (g_β·β_j + g_w·Σ_i w_ij·o_i − g_a·a_j − s_j + g_I·I_j(t)) dt + σ·dW_j
o_j = softmax_hypercolumn(s)
```

- Terms: prior bias β_j (log marginal probability), recurrent evidence Σ w_ij o_i (log pointwise mutual information), leak −s_j, adaptation a_j, external input I_j, neural noise σ·dW_j (temporally correlated OU fluctuations on the support).
- **WTA softmax** within hypercolumn normalizes activities so one unit dominates per step.
- **Adaptation-driven switching**: a_j builds up on the active unit, lowers its support until control passes to the next — adaptation realizes attractor-to-attractor transitions, tracing sequences.
- **Hebbian-Bayes learning**: `w_ij = log(P_ij / (p_i·p_j))` — stores pairwise pointwise mutual information; bias β_j stores each event's log prevalence. Support = accumulated log-evidence of a Naive-Bayes posterior; activity = confidence.

## Ground Truth: Cyclic Markov Chain Environment

- First-order Markov chain combining **probabilistic communities** (dense, frequently visited event groups) with **sparse deterministic chains** (rarely visited).
- Rarity is structural: group size N sets rare-event share (N=5,10,15 → 20%, 10%, 6.7% rare).
- Rare→rare transitions are deterministic, so any departure is signal, not sampling noise.
- Two-level evaluation:
  - **Level 1 (marginal)**: occurrence deviation = rare-event share in replay − ground truth share (signed; positive=over-represented).
  - **Level 2 (conditional)**: KL divergence of the distribution over four transition classes (common→common, common→rare, rare→rare, rare→common).

## Key Mechanism: Two Failure Modes of Deterministic Replay

With σ=0, WTA always selects the argmax-support unit, locking replay into a periodic orbit. Two complementary failures from the same support equation:

1. **Over-representation** (recurrent term run greedily): replay takes the single strongest learned transition each step, skipping common-event alternatives; rare events over-represented *by elimination*. Occurrence deviation > 0.
2. **Under-representation** (bias term overriding recurrent evidence): finite-sample training gives rare targets low prior β_j; a common→rare transition can carry the strongest learned weight (normalized Hebbian rule registers even a single co-occurrence) yet the low prior pulls total support below a common competitor — the transition is **vetoed by low prior**, rare chain never visited. Deviation < 0.

Both failures occur *even though the internal model encodes the rare structure* in its parameters.

## The Fix: Moderate Noise on Supports

Switching on moderate OU noise perturbs supports each step; WTA no longer locks onto the strongest transition but occasionally takes discarded alternatives. **One undirected noise source repairs both opposite distortions**:

- Over-representing instance: replay begins visiting skipped common events; rare excess mass sheds → deviation falls toward 0.
- Under-representing instance: replay leaps past the low-prior barrier into the never-visited rare chain → deviation rises toward 0.

### Inverted-U / Stochastic Resonance Signature

Sweep σ (against training hyperparameter τ_p = plasticity accumulation timescale):

- σ=0: large deviation. Intermediate σ: minimum at σ*≈24 (deviation 0.0143 at 20% rarity). Large σ: degrades again (noise overwhelms learned structure).
- Level-2 KL divergence minimized in the same moderate-σ band — Level-2 tracks Level-1; the optimum is a band, not a single-setting artifact.
- Non-monotonic optimum = classic stochastic resonance: intermediate noise lifts weak (rare) signals across a threshold and ceases to help once it dominates.

## Tolerance Broadening (Robustness Effect)

Sweep σ against four mechanistically distinct parameters (τ_p, prior gain g_β, Bayesian precision g_bayesian, input gain g_I) × three rarities:

- **No parameter setting is accurate at both levels without noise.** With moderate σ, an accurate regime appears spanning a range of parameter values — accuracy becomes insensitive to precise parameter selection.
- Noise doesn't just locate an optimum; it **opens a band** of settings that all support faithful replay. This is the tolerance noise buys.
- Special case g_β: raising it at high noise recovers the marginal but NOT the conditional occurrence — **Level-2 is the stricter test**; prior tuning alone cannot satisfy it.
- Neuromodulatory mapping: DA→(g_β gain + τ_p consolidation timescale), ACh→encode/replay switch (suppresses recurrent evidence g_w, raises afferent g_I, relieves adaptation), NE/arousal→σ.

## Disease Model (Testable Prediction)

Parkinson's disease pushes multiple axes off the tolerant band simultaneously: DA loss weakens prior machinery (g_β, τ_p fall), ACh loss weakens belief precision g_bayesian, aperiodic broadband power rises (σ increases). All three moves drive internal simulation to **over-represent rare events** — matching patient behavior (learning statistics of frequent/rare events yet failing to let priors bias choices) and scopolamine data (cholinergic blockade raises detection at improbable locations). Aging: 1/f slope flattening = drift off the corrective band, mediating working-memory decline.

## Implementation Guide

1. Build the cyclic Markov ground truth: K communities of size N (dense stochastic transitions) + one rare deterministic chain bridging communities; stationary distribution gives rare share ≈ 1/N relative to common.
2. Generate training experience: finite random walks (e.g., 6000 steps); train BCPNN with teacher forcing, noise OFF, local Hebbian-Bayes updates accumulating P_ij, p_i traces with timescale τ_p.
3. Freeze all learned parameters. Remove external input. Replay autonomously with σ>0 OU noise on supports only (not on weights).
4. Evaluate: occurrence deviation (Level 1) and transition-class KL (Level 2) vs ground truth, computed analytically from the chain.
5. Sweep σ × at least one other parameter to map the tolerant band; verify the inverted-U and that the accurate regime is 2D-extended (band), not a point.

## Pitfalls & Best Practices

- **Noise on the wrong variable**: inject noise on unit supports (activation), NOT on synaptic weights — follows BCPNN lineage; supports degrade gracefully.
- **Temporally correlated, not white**: the corrective noise is an OU process (temporally correlated), matching membrane-level integration of multi-source variability.
- **Do not retrain or retune during replay**: the whole point is that correction comes from noise alone with fixed parameters. If you need parameter changes, you've left the scope of this result.
- **Don't claim noise improves learning**: parameters are fixed during replay; results show expression-stage correction only. Noise-assisted replay *could* refine the model if replay engages plasticity later — that's a testable prediction, not a finding.
- **Level-2 KL is the strict test**: reporting only marginal shares can hide broken conditional structure (e.g., g_β tuning case).
- **g_I low-value region is a training failure**, not a robustness effect — set aside in tolerance analysis.

## Extensions & Applications

- Generative replay layers for world models needing rare-event coverage (near-miss/danger events must remain simulable).
- Noise-schedule design for sampling-based recurrent networks (neural-sampling theories: stochastic dynamics visit states ∝ probability).
- Clinical: perturbation experiments on noise band (aging, PD, ataxia channelopathy) as drift-off-optimum accounts; aperiodic activity (1/f slope) as the observable proxy of σ.
- Bridge to stochastic-resonance accounts of tinnitus (brain injects internal noise to recover faint afferents at the cost of phantom percept).
- RL: stochastic-unit recurrent policies transfer better than deterministic ones — moderate internal noise as generative-sampling layer complementing experience-dependent learning.

## Related Skills

- `noise-accelerated-kramers-neural-manifold` — noise-driven escape dynamics on manifolds (Kramers view of the same inverted-U physics)
- `attractor-metadynamics-neural` — slow adaptation reshaping attractor landscapes
- `intrinsic-noise-consolidation-continual-learning` — noise as computational resource in consolidation
- `plasticity-inverse-configurational-constraint` — prospective system properties defined counterfactually (companion theory)

---
arXiv:2609.18033 · q-bio.NC · 16 Sep 2026 · Zhang, Herman, Chao (co-senior: Zhang & Chao)