Many imaging and sensing problems have to rebuild a signal from measurements that throw information away. The best-known case is phase retrieval: a camera or detector records only the intensity of light, never its phase, and the image must be inferred from those intensities alone. The standard iterative solvers for this problem often stall in false solutions. One example is the "twin image," a corrupted mix of the true object and its flipped copy. Getting out of these traps usually means adding something extra: random noise, structured kicks, or careful restarts. All of that adds computation on hardware that is already power-hungry.
Spiking neuromorphic chips work differently. They pass binary pulses ("spikes") between simple neurons and replace multiplications with additions, which is far cheaper in energy. When such a chip represents a number by how often a neuron fires over $T$ time steps, the number comes out noisy, and the noise shrinks as $T$ grows. Precision costs energy, and low precision brings noise with it. This proposal turns that weakness into the escape mechanism. We run an iterative phase-retrieval solver as a spiking network and schedule the number of time steps per iteration: few steps early on, which is cheap, noisy, and good at hopping out of traps, and more steps near convergence, which is expensive but precise. We also drop back to few steps when the spike statistics show the solver has stalled. The claim is that this single control variable is at once the energy budget and the annealing temperature. If so, a spiking solver can escape stagnation while spending less energy than a fixed-precision spiking solver, and much less than a floating-point solver with explicit perturbations.
Core Claim
Rate-coded spiking computation of a matrix-vector product is an unbiased but noisy estimate. Its error variance falls roughly as $1/T$, because a stochastic spike train of length $T$ estimates a firing probability with binomial variance. An iterative projection solver run this way therefore behaves like a noisy fixed-point iteration,
$$x_{k+1} = P(x_k) + \xi_k, \qquad \mathbb{E}[\xi_k]=0,\quad \mathrm{Var}[\xi_k] \approx c/T_k,$$
where $x_k$ is the current estimate, $P$ is one step of a projection algorithm such as Hybrid Input-Output (HIO), $\xi_k$ is the spike-coding error, $T_k$ is the number of time steps used at iteration $k$, and $c$ depends on the measurement operator's norm and the rate-scaling constant. Dynamic energy on event-driven hardware scales with the number of synaptic events, so
$$E_{\text{total}} \approx e_{\text{AC}} \sum_k T_k\, S_k,$$
where $e_{\text{AC}}$ is the energy of one accumulate operation and $S_k$ is the number of synaptic events per time step at iteration $k$. An increasing schedule $T_k$ is therefore simultaneously a cooling schedule and an energy-saving schedule. Escape noise does not cost extra energy. It is the energy you did not spend.
What would be new is not putting phase retrieval on a spiking chip. It is the specific coupling of the substrate's precision–energy tradeoff to landscape escape, plus a stagnation trigger read directly off spike statistics.
Position Relative to Prior Work
The closest method-supplying work is Computing Generalized Matrix Inverse on Spiking Neural Substrate. It maps a recurrent least-squares solver onto IBM TrueNorth using norm-based input scaling and stochastic spike encodings, and reports 10× energy gains over FPGAs and 100× over ARM cores. We reuse its scaling and stochastic-encoding machinery. That work, however, treats a single convex basin and regards spike noise purely as error to be bounded. We move to a non-convex problem where the same noise does useful work.
Two papers motivate the problem from the optimization side. Removal of algorithmic stagnation by augmented iterative phase retrieval (AIPR) shows that HIO stagnates on centrosymmetric supports with a correlation as low as 0.48. Structured operators that add or subtract a 180°-rotated phase raise it above 0.96. AIPR triggers these operators only after a hand-set 800 HIO iterations, with no automatic detection. Estimating Multi-chirp Parameters using Curvature-guided Langevin Monte Carlo escapes many shallow minima with Langevin noise whose scale is set by an expensive Hessian trace. We replace both the hand-set trigger and the curvature computation with a cheap signal: whether the spike-count residual is still falling or only fluctuating.
Navigating Local Minima in Quantized Spiking Neural Networks is the closest work that joins spiking networks and local minima. It escapes minima during weight training with cosine learning-rate restarts. Its landscape is the network's parameter space, not a physical measurement model, and the solver does not run on spiking dynamics at inference time. We borrow its intuition that periodic "reheating" helps, and apply it to the inference-time dynamics themselves.
DeepPhaseCut amortizes phase retrieval into a floating-point feed-forward network. It is a strong accuracy reference, but it is not an iterative solver and says nothing about energy.
One prior result partly constrains the idea. The regularizing effect of spike noise reported in Deep Spiking Neural Networks for Large Vocabulary Automatic Speech Recognition is suggestive, but it concerns trained classifiers, not iterative solvers. The twin-image trap is also a symmetric saddle that roughly isotropic noise may not break. We treat that as an open empirical question, not an assumption.
Research Program
We first build a simulated spiking HIO and a spiking amplitude-flow solver. In both, the Fourier or Gaussian measurement operator is a fixed synaptic matrix, and the current estimate is rate-coded with stochastic spikes over $T_k$ steps. Pointwise steps run as cheap neuron-local operations: the magnitude projection and the support constraint. We also measure the coding noise directly to test whether its variance really scales as $c/T$ and whether it is roughly isotropic or concentrated along particular directions. This determines whether the Langevin analogy holds.
Second, we design schedules. These include monotone increasing $T_k$, and a reheating rule that resets $T_k$ low when a running estimate of residual descent falls below a fraction of its fluctuation. Both quantities come for free from accumulated spike counts at the output neurons.
Third, we test a structured variant. The AIPR operators are a 180° rotation, which is a fixed permutation, plus an add or subtract. On a spiking substrate that is just spike routing with no arithmetic. We test spiking HIO with this cheap structured kick fired by the automatic trigger. This variant exists because isotropic noise alone may fail on twin images.
Energy is estimated by counting accumulates and multiply-accumulates, using the 0.9 pJ versus 4.6 pJ per-operation figures in 45 nm CMOS from SpikingBERT. If hardware access permits, we will also validate on a small instance that fits within one chip's synapse budget. The main claims will rest on simulation with explicit operation counts, and we will state that limitation openly.
Baselines
The naive reference is floating-point HIO with no perturbation, run on the same images, supports, and iteration budget. On the AIPR benchmark it is expected to reproduce stagnation near the reported correlation of 0.48. On CelebA, DeepPhaseCut reports HIO more than 7 dB below its 27.119 dB.
The strongest existing iterative competitor is floating-point HIO plus AIPR. That paper reports a Pearson correlation above 0.96 under ideal conditions and 0.91 under low-light Poisson noise on Cameraman and Peppers. We run it with its published 800-iteration trigger and also with a tuned trigger, giving it the same hyperparameter search budget as our schedules. Beating it means matching its correlation, within run-to-run variation over at least 50 random initializations, at lower estimated total energy. Beating its success rate without a hand-set trigger would be a secondary win.
The strongest amortized reference is DeepPhaseCut on CelebA at 27.119 dB PSNR. We do not expect an iterative spiking solver to beat its accuracy. We report the accuracy–energy point honestly, and also test both methods on out-of-distribution images, where a trained network has no guarantee.
For Gaussian measurements at $n=64$ we use spectral initialization plus Wirtinger flow, and GESPAR. One-Bit Quadratic Compressed Sensing reports GESPAR at an NMSE of $10^{-5}$ and Block SKM at $10^{-7}$. For CG-LMC on multi-chirp mixtures there is no reported result found on Lacuna giving raw error values, so we rerun it ourselves if we include that task.
Two ablations isolate the key ingredient. The first is spiking HIO at fixed high $T$: is the schedule what helps, or just spiking? The second is floating-point HIO with injected Gaussian noise whose variance follows the same $c/T_k$ schedule. That control tells us whether any accuracy gain is specific to spikes. We expect it is not. The claim then reduces to the energy argument, which is still the point.
Why not just…?
Why not just use spectral initialization and Wirtinger flow, which provably avoid spurious minima? The guarantees in Median Truncated Wirtinger Flow and Finding robust minimizer for non-convex phase retrieval rely on random Gaussian-like measurements. Fourier measurements with centrosymmetric supports are exactly where HIO's stagnation is documented. On the Gaussian $n=64$ benchmark this baseline is a real threat, and we expect it to win or tie on accuracy there. The question there is energy only. If a low-precision spiking spectral-plus-flow solver matches it, that is an acceptable result and becomes the recommended design for that regime.
Why not just run floating-point HIO with AIPR on a microcontroller? It works, and it is the accuracy bar. Its weaknesses are the hand-set trigger and full-precision multiply-accumulates for every one of hundreds of iterations. The decisive comparison is the energy-to-reach-0.96-correlation curve, which separates the two directly. If HIO plus AIPR with a tuned trigger is cheaper even under the accumulate energy model, the proposal fails on its central claim.
Why not just use low-bit fixed-point HIO with dithering on ordinary hardware? This is the most serious alternative. Dithered quantization also gives precision-dependent noise and could be scheduled the same way. Our matched-noise floating-point ablation, plus a fixed-point dithered variant counted at multiply-accumulate cost, addresses it. If the dithered variant matches us at comparable energy, the honest conclusion is that the insight is "precision scheduling as annealing" and is hardware-agnostic. Spiking substrates would then be one cheap way to realize it, not a necessity.
Why not just distill DeepPhaseCut into a spiking network? That would cut energy per inference but inherit the trained network's limits. Every input then needs a training set, and there is no mechanism to recover when the network lands in a wrong reflection on unfamiliar objects. Our out-of-distribution test probes this, though we acknowledge it as complementary rather than competing work.
Proposed Experiments
The primary benchmark is Fourier phase retrieval on Cameraman and Peppers with centrosymmetric supports, under ideal and low-light Poisson conditions, following AIPR's protocol. Secondary benchmarks are CelebA Fourier phase retrieval following DeepPhaseCut's protocol, and real Gaussian phase retrieval at $n=64$. For each method we report Pearson correlation or PSNR, the fraction of initializations that escape stagnation, and estimated energy from operation counts. Accuracy is plotted against energy, not iterations.
Three checks are decisive. First, coding-noise variance must scale near $1/T$; otherwise the annealing interpretation is wrong. Second, scheduled-$T$ spiking HIO must escape twin-image stagnation more often than fixed-high-$T$ spiking HIO at lower energy; this is the core claim. Third, the spike-statistic trigger with the routed rotation kick must reach AIPR-level correlation without a hand-set iteration count. A negative result on the second check, with a positive result on the third, would still be informative. It would mean structured kicks are needed and free noise is not enough, which redirects the design toward cheap routed perturbations.
References
- Computing Generalized Matrix Inverse on Spiking Neural Substrate
- Removal of algorithmic stagnation by augmented iterative phase retrieval
- Estimating Multi-chirp Parameters using Curvature-guided Langevin Monte Carlo
- Navigating Local Minima in Quantized Spiking Neural Networks
- DeepPhaseCut: Deep Relaxation in Phase for Unsupervised Fourier Phase Retrieval
- Deep Spiking Neural Networks for Large Vocabulary Automatic Speech Recognition
- SpikingBERT: Distilling BERT to Train Spiking Language Models Using Implicit Differentiation
- One-Bit Quadratic Compressed Sensing: From Sample Abundance to Linear Feasibility
- Provable Non-convex Phase Retrieval with Outliers: Median TruncatedWirtinger Flow
- Finding robust minimizer for non-convex phase retrieval