Pith. sign in

REVIEW 3 major objections 7 minor 15 references

Noise That Helps: Steering Analog Chip Noise to Protect Old Memories

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

2026-07-09 22:50 UTC pith:CS2EL75S

load-bearing objection Genuinely new framing for synaptic consolidation; simulation evidence is solid, silicon result is a proof of concept. the 3 major comments →

arxiv 2607.06924 v1 pith:CS2EL75S submitted 2026-07-08 cs.LG cs.NE

Intrinsic-Noise Consolidation: A Doob-Barrier-Conditioned Diffusion Turns Analog Device Noise into a Continual-Learning Resource

Gunner Levi Howe This is my paper
classification cs.LG cs.NE
keywords noiseintrinsicruleanalogbarrierconsolidationmatchedretention
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Analog neuromorphic chips are intrinsically noisy, and that noise is normally treated as a nuisance that degrades accuracy. This paper proposes that the noise itself can be harnessed to prevent catastrophic forgetting — the overwriting of old memories when a network learns new tasks. The mechanism works by conditioning each synapse's stochastic dynamics on the event that the weight never crosses a barrier around its consolidated value, using a construction from probability theory called a Doob h-transform. The conditioned process acquires an extra restoring drift proportional to the noise variance, so more noise produces a stronger pull back toward the memory — up to the point where the raw diffusion overwhelms the steering. This competition predicts an inverted-U: moderate noise improves retention of previously learned tasks, while too little or too much does not. The paper distinguishes this barrier-conditioning from the plain mean-reverting anchor it also contains (shared with prior methods EWC, OU Adaptation, and MESU), and claims novelty only in the conjunction of the Doob framing and the non-monotone noise-retention signature. A pre-registered experiment on Split-MNIST confirms the inverted-U with a 10.9-point retention lift at an interior noise optimum, while matched anchored-drift controls are monotonically harmed by noise. The effect survives ablation (removing the conditioning flattens the curve), reproduces on a second task stream and under a hardware-faithful noise model, and is demonstrated on real BrainScaleS-2 silicon, where the chip's own intrinsic noise — steered by the barrier conditioning — retains a prior task 15.6 points better than the unconditioned control.

Core claim

Conditioning a synaptic weight's diffusion on never crossing a memory-critical barrier (via a Doob h-transform) produces a restoring drift whose strength scales with the noise variance itself. This creates a regime where increasing intrinsic analog noise improves sequential-task retention up to an interior optimum, then degrades it — an inverted-U that plain anchored-drift consolidation methods cannot produce. The effect is isolated to the barrier conditioning (ablating it removes the inverted-U), and it operates on real silicon where the chip's own noise serves as the diffusion.

What carries the argument

The Doob h-transform applied to per-synapse weight dynamics. Each weight is modeled as a diffusion conditioned on survival within an interval (mu minus b, mu plus b) around its consolidated value. The ground-state survival probability for Brownian motion killed at the interval ends is h(w) = cos(pi(w - mu)/(2b)), giving a steering drift of sigma-squared times the derivative of log h with respect to w. This drift is a restoring force toward the anchor that diverges at the barrier and scales with the noise variance, so the noise itself powers the confinement. The barrier half-width is set per synapse by the Fisher information, so important synapses get tight barriers and unimportant ones get宽松

Load-bearing premise

The steering drift uses the infinite-horizon (ground-state) survival probability for a Brownian particle confined between two absorbing barriers. At the finite timescales relevant to task learning, the true survival probability differs from this ground-state form, and the paper does not test how sensitive the inverted-U is to this approximation. If the finite-horizon correction is large, the steering force would change shape and the optimal noise level could shift or vanish.

What would settle it

Pre-registered as GATE F: if increasing intrinsic noise did not improve retention beyond the unconditioned anchor (i.e., no inverted-U with an interior optimum beating both zero-noise and high-noise ends), the mechanism would reduce to the already-published anchored-drift methods (OUA/MESU/EWC) and the paper would have no contribution. The gate passed: retention rose 10.9 points at an interior noise optimum (paired Wilcoxon p=0.004), while all matched controls were monotonically decreasing in noise.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the inverted-U is robust, analog neuromorphic hardware gains a consolidation advantage that digital accelerators lack: the noise that degrades digital inference becomes a free resource for memory stability on analog substrates.
  • The Doob barrier-conditioning framing could extend to other stochastic dynamical systems where confinement to a safe region is desired — for example, keeping neuronal firing rates within physiological bounds or maintaining robot control parameters within safety margins under actuator noise.
  • The per-synapse barrier width tied to Fisher information suggests a natural connection to geometric approaches to continual learning, where the local loss landscape curvature determines how much each parameter can move before corrupting a prior task.
  • If the finite-horizon correction to the ground-state h-transform is small at learning timescales, the rule could be simplified to a closed-form synaptic update suitable for on-chip plasticity circuits; if it is large, finite-horizon survival functions would need to be approximated, potentially changing the optimal noise level.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The energy argument implies a scaling law: as analog devices shrink and become noisier, the consolidation dividend grows, widening the gap between analog and digital substrates for continual-learning workloads. This could make noisy analog hardware preferable for edge devices that must learn sequentially without replay buffers.
  • The mechanism might interact constructively with stochastic resonance phenomena in biological synapses, suggesting that the brain's noisy analog substrate could be exploiting the same barrier-conditioning principle — though the paper does not make this biological claim.
  • If the barrier width is set by Fisher information and the steering scales with noise variance, there may be an optimal relationship between barrier width and noise level that generalizes across architectures, potentially yielding a design rule for analog continual-learning chips: tune the barrier geometry to the device noise profile.
  • The single-seed on-silicon result (15.6-point retention gain) is a proof of concept at one operating point; if the stability-plasticity trade can be balanced by tuning the barrier coupling, the mechanism might yield net accuracy gains rather than just retention shifts, which would strengthen the practical case considerably.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. This manuscript proposes casting per-synapse memory consolidation as a Doob h-transform: conditioning a synaptic weight's stochastic dynamics on never crossing a memory-critical barrier around its consolidated value. The conditioned diffusion acquires an extra drift term σ²∂_w log h(w), which is a restoring force amplified by the noise variance itself and diverges at the barrier. The paper explicitly disclaims novelty for the anchored drift −s(w−μ) (acknowledging it as a limit of OUA, MESU, and EWC), and claims novelty only in the conjunction of (a) the Doob barrier-conditioning as a synaptic rule and (b) the falsifiable prediction that increasing intrinsic noise non-monotonically improves sequential-task retention (an inverted-U). The prediction was pre-registered as a go/no-go gate and passes on Split-MNIST (8 seeds, 10.9-point lift, p=0.004). The paper includes ablations, a device-faithful BrainScaleS-2 noise emulation, a second task stream, a forward-noise realization, on-silicon noise measurement, and a single-seed on-chip demonstration.

Significance. The paper's intellectual honesty is a genuine strength: it cleanly separates the surrendered drift from the claimed contribution, pre-registered its falsifier with git-verifiable timing, and reports negative results openly (Benna–Fusi failing to learn, replay beating the rule). The Doob h-transform framing for synaptic consolidation is, to my knowledge, genuinely novel. The inverted-U signature is a crisp, falsifiable prediction that anchored-drift methods cannot produce by construction, and the experimental isolation (κ ablation, barrier-tracking) is well-designed. The on-silicon measurement and hardware-in-the-loop demonstration, even at single seed, add credibility to the practical motivation.

major comments (3)
  1. §3 (end of Method) and §6: The finite-force cap on the Doob steering drift is a load-bearing element whose interaction with σ² is not adequately isolated from the claimed mechanism. The uncapped h-transformed process is exactly conditioned on never crossing the barrier; the capped process is not. The cap replaces the σ²-scaled divergent tan force with a σ²-scaled constant force whenever w is near the barrier. The paper acknowledges (§6) that the cap 'removes the noise (σ²) amplification of the steering where it binds' on 8.8% of Doob steps at the operating point. Since the σ²-scaling of the steering is the entire proposed mechanism, the inverted-U could be partly driven by the cap–σ² interaction rather than by pure Doob conditioning. A sensitivity analysis over the cap fraction (or an alternative regularization of the singular drift, e.g., soft-clipping) would determine whether the cap's
  2. §3, Eq. (2) and §6: The ground-state h-transform h(w) = cos(π(w−μ)/(2b)) is the infinite-horizon quasi-stationary survival probability for Brownian motion killed at interval ends. The paper acknowledges using this rather than a finite-horizon survival function, but does not test sensitivity to this approximation. If the finite-horizon correction is substantial at the timescales relevant to task learning (which may be far from quasi-stationary), the steering drift σ²∂_w log h would differ from the ground-state form, potentially altering the inverted-U. A brief analysis or simulation comparing ground-state vs. finite-horizon h would clarify whether this approximation is benign or distorting.
  3. §3: The cap fraction value and the barrier scale b_0 are free parameters, but the paper does not report how they were selected or whether they were part of the pre-registration. Since the optimum σ* tracks b_0 (Fig. 2b), and the cap interacts with σ² (see first major comment), the choice of these parameters is not obviously neutral. The paper should state whether b_0 and the cap fraction were fixed before the pre-registered experiments or tuned post hoc, and should report sensitivity to the cap fraction (at minimum: does the inverted-U survive when the cap is tightened or loosened by an order of magnitude?).
minor comments (7)
  1. §4.4, Fig. 4a: The Benna–Fusi cascade baseline did not learn (50.6%, chance). The paper flags this honestly, but it would strengthen the comparison to note whether this is a known failure mode of the cascade on small MLPs or a tuning failure on the authors' part.
  2. §4.8: The on-silicon result reports task-2 plasticity of 46.6% vs. 64.0% for the control. The abstract says 'matched average accuracy,' but the body says 'essentially matched' (58.1% vs. 59.0%). The abstract's phrasing is slightly stronger than the body's; aligning them would help.
  3. Fig. 1: The y-axis label 'retention (mean past-task acc.)' could be clearer about whether this is accuracy averaged over all past tasks at the end of the full sequence (i.e., final Split-MNIST after 5 tasks).
  4. §4.6, Fig. 7b: The num_sends exponent is fitted as −0.47 from 4 data points. Stating the R² or confidence interval on this fit would help readers assess whether 1/√N is a reasonable model.
  5. §2: The related-work section is thorough but dense. A brief table summarizing which existing methods use barriers, h-transforms, noise optima, and hardware would make the novelty boundary crisper.
  6. §4.7, Fig. 8b: The three coupling values shown all yield σ* ≈ 0.05–0.12, but the text says the optimum moves 'down to 0.05.' Clarifying whether 0.05 is the lowest achievable or just one of three points would help.
  7. The energy model (§4.4) is described as 'not robust to its constants' (range 4%–41%). The 23% figure is reported as 'one modelled point.' This is appropriately caveated, but the figure (Fig. 4b) may overstate precision; a shaded range rather than a point would better convey the uncertainty.

Circularity Check

0 steps flagged

No circularity found: the derivation chain is self-contained, the inverted-U prediction is not forced by construction, and the h-transform math is standard.

full rationale

The paper's derivation chain is clean and non-circular. (1) The Doob h-transform and the ground-state survival probability h(w) = cos(π(w-μ)/(2b)) for killed Brownian motion is a standard mathematical result, not fitted to data or defined in terms of the prediction. (2) The σ²-scaling of the steering drift (σ²∂_w log h) is a direct mathematical consequence of the h-transform formula — it is true by construction that more noise amplifies the steering force. However, the *inverted-U prediction for retention* is NOT tautological: it requires the empirical claim that at moderate σ the σ²-steering wins (good retention) while at high σ the raw diffusion overwhelms it (bad retention). This competition between two σ-dependent terms with different scaling (σ² vs σ) is a genuine prediction, which is why the authors pre-registered it as a falsifiable gate. (3) The barrier scale b_i = b_0/√(1 + s_i/median(s)) uses Fisher information — a principled choice, not fitted to retention outcomes. (4) The paper is single-authored with no self-citation chain; the anchored drift is explicitly surrendered as a known limit of OUA/MESU/EWC (external works). (5) The claim that anchored-drift methods 'cannot produce' the inverted-U is a mathematical distinction: their drift lacks σ²-scaling, so retention is necessarily monotone-decreasing in noise — not a circular assertion. The skeptic's concern about the finite-force cap confounding the σ²-scaling is a validity/correctness concern (does the experiment test the theory?), not a circularity concern (is the theory equivalent to its inputs?). No step in the derivation reduces to its own inputs by definition or self-citation.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The axiom ledger is lean. The core mathematical ingredients (Doob h-transform, killed Brownian motion) are standard results, not invented. The main domain assumptions (additive noise, infinite-horizon approximation, diagonal Fisher sufficiency) are either measured on silicon or acknowledged as limitations. The one ad hoc element (finite-force cap) is disclosed but not sensitivity-tested. No new physical entities are postulated; the barrier and steering drift are mathematical constructs with falsifiable handles.

free parameters (5)
  • b_0 = 0.1, 0.2, 0.4 (swept)
    Global barrier scale parameter controlling the half-width of the memory-critical interval. Swept in Fig. 2b to show the optimum tracks the barrier. Not fitted to the retention data but chosen by hand.
  • sigma = 0.02 (optimum)
    Intrinsic noise amplitude. The inverted-U optimum sigma* is the key prediction, not a fitted parameter per se, but the operating point was calibrated before the gated seeds.
  • kappa = 1.0 (full conditioning)
    Steering strength parameter interpolating between unconditioned OU (kappa=0) and full h-transform (kappa=1). Swept in Fig. 2a as an ablation.
  • s_i (Fisher importance) = online-EWC running sum
    Per-synapse importance from online EWC. Standard from prior literature, not invented here.
  • Doob move cap fraction = unspecified fraction of b_i
    Finite-bandwidth cap on the per-step Doob move to stabilize the singular drift. The paper states it caps at a fraction of b_i but does not specify the fraction. This is a tuning parameter.
axioms (5)
  • standard math Ground-state h-transform: h(w) = cos(pi(w-mu)/(2b)) is the survival probability for Brownian motion killed at interval ends.
    Standard result from stochastic process theory (e.g., Doob h-transform, quasi-stationary distributions). Invoked in Eq. 2, Section 3.
  • domain assumption The infinite-horizon (ground-state) approximation is valid at the timescales of task learning.
    The paper uses the ground-state h-transform rather than a finite-horizon survival function. Acknowledged as a limitation in Section 6 but not tested for sensitivity.
  • domain assumption Intrinsic device noise is additive and trial-to-trial independent at the update timescale.
    Measured on BSS-2 silicon (Section 4.6) and found to hold: 94.6% additive, trial-to-trial independent. This is the noise class the mechanism requires.
  • domain assumption The diagonal Fisher information is a sufficient importance metric for setting per-synapse barriers.
    The barrier half-width b_i = b_0/sqrt(1+s_i/median(s)) uses the diagonal Fisher. Standard in EWC literature but an approximation.
  • ad hoc to paper The finite-force cap on the Doob move does not qualitatively alter the mechanism.
    The cap removes the sigma^2 amplification where it binds (8.8% of Doob steps at the operating point, per Section 6). The paper does not test whether removing the cap or changing its value affects the inverted-U.
invented entities (2)
  • Memory-critical barrier (mu +/- b) independent evidence
    purpose: Defines the safe region for each weight; crossing it means memory loss.
    The barrier is read off the Fisher metric (Section 3) and its scale is swept experimentally (Fig. 2b), showing the optimum tracks it. Falsifiable: if the barrier geometry did not set the operating point, the prediction would fail.
  • Doob steering drift (sigma^2 * d/dw log h) independent evidence
    purpose: The noise-amplified, barrier-divergent restoring force that is the paper's core mechanism.
    Derived from standard mathematics (not invented), but its application as a synaptic rule is new. The ablation (kappa: 1->0, Fig. 2a) provides falsifiable evidence that it is the conditioning, not generic noise, producing the inverted-U.

pith-pipeline@v1.1.0-glm · 16388 in / 3178 out tokens · 419355 ms · 2026-07-09T22:50:05.267059+00:00 · methodology

0 comments
read the original abstract

On analog neuromorphic hardware, intrinsic device noise is normally an accuracy tax. We ask whether it can instead consolidate memories. We 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 conditioned diffusion gains an extra drift sigma^2 d/dw log h, a restoring force amplified by the noise variance itself that diverges at the barrier. We are explicit about novelty: the anchored drift -s(w-mu) our rule also contains is not ours (the limit of OUA, MESU, and EWC), and we surrender it. We claim only the conjunction of (a) the Doob barrier-conditioning as a synaptic rule, to our knowledge unclaimed (every h-transform use we found is generative modeling, none synaptic), and (b) a falsifiable prediction: increasing intrinsic noise non-monotonically improves sequential-task retention, an inverted-U that anchored-drift methods cannot produce. We pre-registered this as a go/no-go gate; it passes. On single-head Split-MNIST (8 seeds) the rule lifts retention 10.9 points at an interior optimum (paired Wilcoxon p=0.004), while matched OU/EWC/MESU anchors are monotone. Ablating the conditioning removes the effect; the optimum tracks the barrier; the inverted-U survives a second task stream and the realization where noise enters the forward pass. We then measure the intrinsic noise on real BrainScaleS-2 silicon (additive, trial-to-trial independent, tunable via on-chip averaging) and run the rule on the chip with its noise in the training loop: barrier-conditioning retains a prior task 15.6 points better than the matched control at matched average accuracy, a stability-plasticity shift, not a net-accuracy win (single seed; retention measured, energy modelled). Intrinsic analog noise thus becomes a consolidation dividend a digital accelerator must spend energy to generate.

Figures

Figures reproduced from arXiv: 2607.06924 by Gunner Levi Howe.

Figure 1
Figure 1. Figure 1: GATE F (Split-MNIST, 8 seeds, mean±sd). The barrier-conditioned rule (Doob, red) rises from 57.4% at zero noise to 68.3% at an interior optimum σ ∗ = 0.02 (a 10.9-point lift; paired Wilcoxon p = 0.004 vs. zero, p = 0.004 vs. the high-noise end), then falls. The matched OU, EWC, MESU and naive anchors are monotone-decreasing in noise. Noise is a resource only when steered by the barrier. to 0.0 points (κ=0)… view at source ↗
Figure 2
Figure 2. Figure 2: Isolation. (a) Interpolating the steering strength κ from 0 (unconditioned OU) to 1 (full h￾transform): the inverted-U emerges with the conditioning (lift at κ=0 is 0.0 pts; at κ=1, 13.0 pts). (b) Varying the barrier scale b0: the retention optimum σ ∗ tracks the barrier, evidence the optimum is set by the conditioning geometry, not a generic noise sweet-spot. a modelled 23% fraction of each consolidation … view at source ↗
Figure 3
Figure 3. Figure 3: Device-faithful noise (emulation, not silicon). (a) With the diffusion replaced by the BSS-2 model (colored + multiplicative + fixed-pattern + 6-bit), the inverted-U survives (lift 11.0 pts at σ ∗ = 0.005). (b) Scanning the temporal color ρ of the device noise maps where the mechanism holds — the concrete content of kill condition K2, to be settled on silicon. from 4 points; Fig. 7b), so it is a genuine on… view at source ↗
Figure 4
Figure 4. Figure 4: Baselines (Split-MNIST, 8 seeds). (a) At its noise optimum the barrier-conditioned rule (Doob∗ , 68.3%) is the strongest rehearsal-free method: it matches MESU (65.2%) and significantly beats OU, EWC and naive (our Benna–Fusi did not learn in this configuration, 50.6% ≈ chance — flagged, not counted). Plain replay stores exemplars and scores higher (83.0%) — a different, memory-based budget class, shown fo… view at source ↗
Figure 5
Figure 5. Figure 5: Second modality. (a) On continual Yin-Yang the inverted-U reproduces (lift 7.4 pts at σ ∗ = 0.035); the OU control is flat. (b) As tasks become less similar (larger per-task rotation, more interference) the noise optimum shifts, mapping the mechanism’s operating regime. is essentially matched (ours 58.1% vs. the control’s 59.0%) while retention strongly favours ours. This is a stability-plasticity shift, n… view at source ↗
Figure 6
Figure 6. Figure 6: Mechanism. (a) The Doob steering drift σ 2∂w log h for several noise levels: a barrier-divergent restoring force whose strength scales with the noise variance. (b) The retention optimum coincides with a forgetting minimum (1−forgetting peak) at σ ∗ ; plasticity is roughly noise-insensitive. The inverted-U is a protection effect, not a plasticity tradeoff. 6 Limitations Small single-head MLP testbeds; a dia… view at source ↗
Figure 7
Figure 7. Figure 7: Intrinsic noise measured on BrainScaleS-2 silicon (chip hxcube7fpga3chip61 1). (a) The trial-to-trial noise std is flat in the signal (additive, 94.6%); the CV therefore falls as the signal grows. (b) num sends averages the relative noise as ≈ 1/ √ N (measured exponent 0.47) — the on-chip effective-noise knob. in the hardware-faithful forward-noise realization, is isolated to the barrier conditioning, and … view at source ↗
Figure 8
Figure 8. Figure 8: Forward-noise (hardware-faithful) realization. (a) With noise in the forward pass, Doob is an inverted-U (lift 12.9 pts, p = 0.004) and OU is flat; without the importance clamp both collapse to chance (dotted). (b) The Doob-steering coupling tunes the retention optimum down to σ ∗ = 0.05 — into the device-reachable noise band (shaded) — so the mechanism is portable to a given chip’s intrinsic-noise level. … view at source ↗
Figure 9
Figure 9. Figure 9: On real BrainScaleS-2 silicon (hardware-in-the-loop; the chip’s intrinsic noise is the diffusion). After learning task 2, the barrier-conditioned rule retains task 1 at 69.6% vs. 54.0% for the matched uncon￾ditioned control (+15.6 pts). Single seed, one operating point; the rule trades some task-2 plasticity for the retention (stated in text). erated neuromorphic system with hybrid plasticity. Frontiers in… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages · 10 internal anchors

  1. [1]

    Djohan Bonnet, Kellian Cottart, Tifenn Hirtzlin, Tarcisius Januel, Thomas Dalgaty, Elisa Vianello, and Damien Querlioz

    doi: 10.1038/nn.4401. Djohan Bonnet, Kellian Cottart, Tifenn Hirtzlin, Tarcisius Januel, Thomas Dalgaty, Elisa Vianello, and Damien Querlioz. Bayesian continual learning and forgetting in neural networks.Nature Communications, 16,

    work page doi:10.1038/nn.4401

  2. [2]

    Bayesian continual learning and forgetting in neural networks

    doi: 10.1038/s41467-025-64601-w. arXiv:2504.13569; introduces MESU (Metaplasticity from Synaptic Uncertainty), Eq

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1038/s41467-025-64601-w

  3. [3]

    Reflected Schr\"odinger Bridge for Constrained Generative Modeling

    arXiv:2401.03228. Yuanqi Du, Michael Plainer, Rob Brekelmans, Chenru Duan, Frank No´ e, Carla P. Gomes, Al´ an Aspuru- Guzik, and Kirill Neklyudov. Doob’s lagrangian: A sample-efficient variational approach to transition path sampling. InInternational Conference on Learning Representations (ICLR),

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    arXiv:2410.07974. 12 0.0 0.5 1.0 1.5 2.0 forward (activation / MAC) noise 0.50 0.55 0.60 0.65 0.70 0.75 0.80retention (a) mechanism survives forward noise (the hardware-relevant case) Doob (ours) OU (control) Doob, no clamp (collapses) 0.0 0.1 0.2 0.3 forward noise (device-reachable band shaded) 0.64 0.66 0.68 0.70 0.72 0.74 0.76 0.78retention (b) couplin...

    work page internal anchor Pith review Pith/arXiv arXiv

  5. [5]

    Ornstein-Uhlenbeck Adaptation as a Mechanism for Learning in Brains and Machines

    doi: 10.3390/e26121125. arXiv:2410.13563. Jeremy Heng, Valentin De Bortoli, Arnaud Doucet, and James Thornton. Simulating diffusion bridges with score matching.arXiv preprint arXiv:2111.07243,

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.3390/e26121125

  6. [6]

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A

    doi: 10.1109/ISSCC.2014.6757323. James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National ...

    work page doi:10.1109/isscc.2014.6757323 2014

  7. [7]

    Rusu and Kieran Milan and John Quan and Tiago Ramalho and Agnieszka Grabska-Barwinska and Demis Hassabis and Claudia Clopath and Dharshan Kumaran and Raia Hadsell , title =

    doi: 10.1073/pnas.1611835114. Ivan Kolesnikov and Nadezhda Semenova. Internal noise in hardware deep and recurrent neural networks helps with learning.arXiv preprint arXiv:2504.13778,

    work page doi:10.1073/pnas.1611835114

  8. [8]

    The Yin-Yang dataset

    arXiv:2102.08211. Luca Manneschi, Ian T. Vidamour, Kilian D. Stenning, et al. Noise-aware training of neuromorphic dynamic device networks.Nature Communications, 16,

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    Noise-Aware Training of Neuromorphic Dynamic Device Networks

    doi: 10.1038/s41467-025-64232-1. arXiv:2401.07387. Toan Nguyen, Kien Do, Duc Kieu, and Thin Nguyen. h-edit: Effective and flexible diffusion-based editing via doob’s h-transform. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1038/s41467-025-64232-1

  10. [10]

    h-Edit: Effective and Flexible Diffusion-Based Editing via Doob's h-Transform

    arXiv:2503.02187. Christian Pehle, Sebastian Billaudelle, Benjamin Cramer, Jakob Kaiser, Korbinian Schreiber, Yannik Strad- mann, Johannes Weis, Aron Leibfried, Eric M¨ uller, and Johannes Schemmel. The brainscales-2 accel- 13 task0 (retain) task1 (learn) 0.0 0.2 0.4 0.6 0.8 1.0accuracy +15.6 pts On real BrainScaleS-2 silicon (hardware-in-the-loop, chip n...

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    Nimrod Shaham, Jay Chandra, Gabriel Kreiman, and Haim Sompolinsky

    doi: 10.3389/fnins.2022.795876. Nimrod Shaham, Jay Chandra, Gabriel Kreiman, and Haim Sompolinsky. Stochastic consolidation of lifelong memory.Scientific Reports, 12:13107,

    work page doi:10.3389/fnins.2022.795876 2022

  12. [12]

    doi: 10.1038/s41598-022-16407-9. Johannes Weis, Philipp Spilger, Sebastian Billaudelle, Yannik Stradmann, Arne Emmel, Eric M¨ uller, Oliver Breitwieser, Andreas Gr¨ ubl, Joscha Ilmberger, Vitali Karasenko, Mitja Kleider, Korbinian Schreiber, Christian Weilbach, and Johannes Schemmel. Inference with artificial neural networks on analog neu- romorphic hardw...

    work page doi:10.1038/s41598-022-16407-9

  13. [13]

    Inference with Artificial Neural Networks on Analog Neuromorphic Hardware

    arXiv:2006.13177. Zeke Xie, Fengxiang He, Shaopeng Fu, Issei Sato, Dacheng Tao, and Masashi Sugiyama. Artificial neural variability for deep learning: On overfitting, noise memorization, and catastrophic forgetting.Neural Computation,

    work page internal anchor Pith review Pith/arXiv arXiv 2006

  14. [14]

    Artificial Neural Variability for Deep Learning: On Overfitting, Noise Memorization, and Catastrophic Forgetting

    arXiv:2011.06220 (2020); introduces Artificial Neural Variability (ANV). Fatima Tuz Zohora, Vedant Karia, Nicholas Soures, and Dhireesha Kudithipudi. Probabilistic metaplasticity for continual learning with memristors.Scientific Reports, 14,

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  15. [15]

    Probabilistic Metaplasticity for Continual Learning with Memristors

    doi: 10.1038/s41598-024-78290-w. arXiv:2403.08718. 14

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1038/s41598-024-78290-w