Some analog chips are noisy: their components flip, drift and fluctuate at random. A line of hardware work treats that noise as a free source of randomness for optimization. The idea is that random jolts help a search escape poor solutions it would otherwise get stuck in, without paying for a digital random-number generator. In treatment-resistant depression (depression that has not responded to standard antidepressants), a growing option is electrical brain stimulation aimed at several interacting regions at once. Choosing how to drive current through a cap of 64 scalp electrodes so that the right regions get the right dose is a hard search problem that is known to get stuck. It gets much harder in the version clinics actually face, where only a handful of electrodes can be active.
This note proposes a deliberately falsifiable test of whether noise-driven analog solvers earn a place in that pipeline. We would cast montage design (the choice of which electrodes are on, with which polarity) as a problem on binary variables. We would generate a benchmark of such problems from depression-relevant brain targets and a realistic head model. Each problem is small enough to solve exactly, so we know the true best answer. We would then measure how close device-noise annealers come to that answer, and at what energy cost, compared with ordinary digital methods. The likely outcome is mixed, and the value of the work is that it settles the question with ground truth rather than asserting it.
Core Claim
Our claim is that multi-target, sparse-electrode stimulation design for depression is a real combinatorial problem of moderate size, roughly 128 binary variables. At that size, three things can be measured cleanly against exact optima. The first is whether hardware-style annealing with realistic device noise finds better montages than convex relaxation followed by rounding. The second is whether the device's own noise matters, or whether the same analog dynamics with the noise switched off do just as well. The third is whether any advantage survives the cheapest alternative of all, a precomputed lookup table.
The hardware literature has shown noise helping on abstract benchmarks. BRIM builds an analog circuit whose voltages settle into low-energy states of an Ising model, a network of binary "spins" with pairwise couplings, and it uses random spin flips to escape local minima on Max-Cut. In-memory factorization uses phase-change-memory switching noise to break the endless loops that trap a deterministic search procedure. Ding et al. survey memristor noise used as a perturbation for annealing. None of these was tested on a problem from neuromodulation, and none compared against an exact optimum on a clinically shaped instance.
The Problem, Stated Precisely
The stimulation literature motivates the problem. Xiao et al. steer current from a 64-electrode cap through a five-layer finite-element head model, a simulation of how current spreads through scalp, skull, cerebrospinal fluid and brain. They report that multi-region optimization falls into local optima and that standard methods starve small or deep targets. They had to restart their software from random starting points to approximate a global solution, and they report no numerical error values. We would replace their healthy working-memory targets with targets taken from the depression literature. These are the dorsolateral prefrontal and orbitofrontal cortex used in the multi-region transcranial direct current stimulation (tDCS) work of Shi et al., and the medial prefrontal and posterior cingulate nodes whose top-down coupling predicts non-response in Zheng et al..
Each electrode $e$ carries two binary variables, $s_e^+$ and $s_e^-$, which mark whether it is an anode or a cathode at a fixed current $I_0$. Its current is therefore $I_e = I_0(s_e^+ - s_e^-)$. The head model gives a linear map from currents to field. Let $a_r$ be the vector of field contributions at target region $r$, $t_r$ the desired field there, $w_r$ an equalizing weight in the spirit of Xiao's weighted schemes, $B$ the matrix of field contributions at non-target tissue, and $K$ the allowed number of active electrodes. The objective is
$$E(s) = \sum_r w_r \left(a_r^\top I(s) - t_r\right)^2 + \lambda \lVert B\, I(s)\rVert^2 + \mu\Big(\sum_e (s_e^+ + s_e^-) - K\Big)^2 + \nu \sum_e s_e^+ s_e^-,$$
where $\lambda$, $\mu$ and $\nu$ weight off-target leakage, the electrode budget and a penalty against an electrode being both anode and cathode. A full current-balance constraint would add one more quadratic penalty. Every term is quadratic in binary variables, so the problem maps directly onto an Ising machine's couplings. The electrode budget is what makes the problem combinatorial rather than convex.
Research Program
We would build three things.
The first is a benchmark generator. From the head model it produces families of instances that vary target sets, target weights, electrode budget $K$ between 4 and 12, and penalty strengths. Every instance is solved to proven optimality offline with a mixed-integer quadratic solver; at 64 electrodes this should be tractable. Every heuristic can then be scored as a gap to the true optimum rather than against another heuristic.
The second is a set of device-realistic annealer emulators. One is a BRIM-style coupled-bistable model following Afoakwa et al.. The others are a memristor-perturbed and a phase-change-memory-perturbed Hopfield-style annealer, whose noise statistics follow the categories in Ding et al. and the switching noise in Langenegger et al.. Each emulator has explicit knobs for the open problems the file flags: amplitude miscalibration, slow drift between runs, and chip-to-chip variation. This lets us test whether uncontrolled noise helps or hurts.
The third is an energy accounting. It compares joules per solved instance across the emulated hardware and a microcontroller-class digital implementation. We would use the roughly 150 µW implant processor of Wang et al. as the reference for what a tight on-device budget looks like. That work is the closest existing hardware-algorithm co-design for neuromodulation. It is fully deterministic, demonstrated on seizure detection, and solves no optimization problem, which is exactly the slot this study probes.
We would state the clinical limit plainly. The targets come from group-level findings, so nothing here shows that a better-scoring montage relieves depression. The contribution is an engineering verdict on one computational step.
Baselines
The naive reference is Xiao et al.'s unweighted least-squares montage, thresholded to the $K$ largest electrodes. It represents what one gets with no combinatorial effort. Its reported failure mode is that one region absorbs nearly all the current; Xiao et al. report the maximum-field variant "almost entirely ignored" a secondary target. No reported numerical error is found on Lacuna for this setting, so we would measure it ourselves on every instance.
The strongest existing method is convex relaxation plus rounding. Here the binary variables are relaxed to continuous currents with an L1 sparsity penalty using Xiao's weighted least-squares formulation, the result is rounded to $K$ electrodes, and a local search polishes it. It gets the same instances, the same head model, and the same wall-clock and tuning budget as the annealers, with sparsity and weights tuned on held-out instances. Beating it means a statistically reliable lower median optimality gap, or the same gap at lower energy, across instance families. A win confined to a few hand-picked instances does not count.
The digital stochastic baseline is software simulated annealing with a standard pseudo-random generator. Its temperature schedule is tuned with the same budget as the hardware emulators. BRIM reports its analog solver as 7 to 8 orders of magnitude faster than software simulated annealing on 2,000-node Max-Cut, settling in 0.25 µs at about 250 mW. That figure comes from a much larger, different problem, so we report time-to-target and energy-to-target on our instances instead of assuming it transfers.
The key ablation is the same analog dynamics with noise removed. For the BRIM-style emulator this keeps only coupling annealing and drops the stochastic flips; for the memory-based emulators it uses an ideal deterministic crossbar. If this ablation matches the noisy version, the noise is not doing the work. The exact mixed-integer solver supplies the optimum that everything is scored against.
Why not just…?
Why not just solve it exactly offline? For a fixed prescription this is probably the right answer, and the benchmark will show how long exact solving takes. A hardware solver is only interesting if montages must be re-solved often on a battery-powered device. An example is a closed-loop stimulator that shifts target weights as a decoded state changes, using a wearable prefrontal EEG signal like the 3-channel recordings in MODMA. If exact solves take seconds on a phone, the hardware case collapses for transcranial use, and we would report that.
Why not just precompute a lookup table? This is the most serious threat. With two or three targets, the space of weight vectors can be gridded and every cell solved in advance. We would measure table size and interpolation error as target count and grid resolution grow, and find where, if anywhere, on-line solving beats storage. If a table under a few hundred kilobytes, comparable to the 160 KB on-chip memory in Wang et al., covers clinically plausible weightings within 1% of optimum, the proposal's practical case fails for this problem.
Why not just use a digital random-number generator? The file claims pseudo-random generation is a power and latency bottleneck, citing Mostafa et al. and the 150× energy figure from Ding et al.. Those claims come from workloads drawing vast numbers of samples. A 128-variable anneal needs far fewer, so the generator's cost may be negligible next to the stimulator's own output power. Our energy accounting compares joules per solve directly instead of inheriting the claim.
Decisive Checks
We would first validate the BRIM-style emulator on Gset instances G22 and G39 to confirm it reproduces qualitatively the behavior Afoakwa et al. report. We would then run all methods on the depression-target benchmark and report four quantities per instance family: optimality gap to the exact solution, the worst-served target's achieved field as a fraction of its desired field (which directly measures Xiao's starvation failure), off-target energy, and energy-to-solution.
The proposal's hypothesis holds only if three things are true together. Noisy annealers beat tuned convex rounding on the gap or on energy at equal gap. They beat their own noise-free ablation. And the advantage persists under the miscalibration and drift sweeps that mimic real devices.
If the noise helps only when perfectly calibrated, that is a useful negative result. It would mean the open controllability problem flagged by Ding et al. is the binding constraint for clinical use, and it tells hardware groups what to fix before targeting neuromodulation. If convex rounding or a lookup table matches everything, we would publish the benchmark and the exact optima, so that later claims about analog solvers for stimulation have ground truth to meet.
References
- BRIM: Bistable Resistively-Coupled Ising Machine
- In-memory factorization of holographic perceptual representations
- Transforming memristor noises into computational innovations
- Multi-Region Brain Stimulation Optimization Using Transcranial Direct Current Stimulation
- Effects of Transcranial Direct Current Stimulation Targeting Dorsolateral Prefrontal Cortex and Orbitofrontal Cortex on Somatic Symptoms in Patients With Major Depressive Disorder: A Randomized, Double‐Blind, Controlled Clinical Trial
- Beyond Depression Symptoms: The Default Mode Network as a Predictor of Antidepressant Response
- Precise and low-power closed-loop neuromodulation through algorithm-integrated circuit co-design
- A multi-modal open dataset for mental-disorder analysis
- Recurrent networks of coupled Winner-Take-All oscillators for solving constraint satisfaction problems