Photonic accelerators do the multiply-accumulate steps of a neural network with light passing through tiny optical rings and interferometers. They are extremely sensitive to temperature. Silicon's refractive index rises with heat, so every ring's resonance drifts, and every encoded weight drifts with it. One prior study reports that a shift of only 0.05 K degrades classification accuracy from above 90% to about 57% (DOCTOR). Current fixes either burn power in microheaters, or periodically pause the computation to probe the hardware and recalibrate it. Precision physics has a different playbook for drifts it cannot remove. It measures the drift with a channel that shares the error but not the signal, divides it out, and then puts an explicit error bar on whatever remains. That leftover is called a residual systematic uncertainty. The reactor experiments that take spectral ratios between two detectors and the Daya Bay two-channel cross-check are the canonical examples.
This note proposes to bring that playbook onto the chip. Each tile of rings gets a small number of reference rings. They sit in the same thermal neighborhood as the weight rings and run on a spare wavelength with a known, fixed weight, so they are read out continuously alongside live inference. Subtracting their drift removes the shared, common-mode part of the thermal shift without interrupting computation. Disagreement among the references then gives a running estimate of the residual, non-common part that cancellation leaves behind. That estimate is used in two ways. It sets the width of calibrated prediction sets, meaning sets of candidate labels guaranteed to contain the true one at a chosen rate. It also decides, on statistical grounds rather than a fixed threshold, when the expensive pause-and-probe recalibration is actually needed.
Core Claim
Most thermo-optic drift within a tile is spatially smooth. A reference ring at distance $d$ from a weight ring should therefore cancel the common shift and leave a residual roughly proportional to the local temperature gradient times $d$; this is a first-order Taylor argument. The residual can be bounded online from the spread among $m$ reference rings per tile, which plays the role Daya Bay's hydrogen-capture channel played against its gadolinium channel: an independent view whose agreement certifies the main result (Daya Bay nH).
Write $\hat r_t$ for the reference-spread estimate of the residual phase or wavelength error at time $t$. The claim has two parts. First, accuracy after cancellation stays near the drift-free level across the drift scenarios DOCTOR uses, with far fewer probe interruptions. Second, prediction sets whose threshold is indexed by binned $\hat r_t$ keep nominal coverage, where both a single global threshold and an online-updated threshold fail during fast hotspot transients.
The new element is not the reference ring itself. Wavelength-locking to a reference is standard practice. What is new is treating the post-cancellation residual as a measured systematic with a calibrated bound, and showing that this bound, rather than a raw temperature reading, is the right quantity to drive both uncertainty reporting and recalibration.
Relation to Prior Work
Three photonic papers motivate the gap.
DOCTOR probes cores with identity inputs, corrects salient blocks by regression, and remaps critical layers to cool tiles. Its controller fires on sensor and error thresholds and costs 0.14%–5% of cycles. We keep its probing as the corrective step but replace its trigger and remove most of its invocations.
SCATTER avoids phase sensitivity with phase-agnostic cores and reroutes light. It reports 91.31% on CNN-FMNIST under thermal variations, but evaluates static thermal maps and gives no confidence bounds.
The modular interferometer accelerator GOA retains coherent phase computation and documents that thermal tuning power can cancel its energy gains. These two papers embody the tension we address: reference cancellation is a way to keep phase-sensitive devices usable without paying for heater locking.
Two physics papers supply the method. Reactor spectral ratios cancel flux and efficiency errors by comparing two baselines. Their stated limitation constrains us directly: cancellation breaks when the two detectors are not identical. That limitation is exactly why we measure the residual instead of assuming it is zero.
NAPS builds prediction sets valid across a continuous nuisance parameter. It keeps 90% coverage where standard conformal sets fall to 40%–60%. It takes seconds per set and scales quadratically in class count, so we cannot run it at inference speed. We borrow its principle of guaranteeing coverage across the nuisance range, and amortize it into a precomputed lookup of thresholds indexed by the binned residual estimate.
Architecture-level photonic modeling is the constraint on the energy accounting. It shows that converters and DRAM dominate energy, so the extra photodetectors and analog-to-digital converters for the reference rings must be charged against the method.
Research Program
We will build a physics-level simulator of microring weight banks. It will model per-ring thermo-optic resonance shift, a time-varying temperature field, and fabrication offsets among rings. The thermal field will include the linear-rise and corner-hotspot drift scenarios DOCTOR evaluates, plus a sharper-gradient regime that is designed to break common-mode cancellation.
Each tile receives $m \in \{1,2,4\}$ reference rings at a controlled distance $d$ from the weight rings. Each reference ring runs on a spare wavelength-division channel, meaning a spare laser wavelength on the same waveguide, with a fixed known weight. Its readout yields the common-mode shift, which is subtracted digitally from the weight commands of that tile. The spread among the $m$ references, after a one-time fit at the factory to remove static fabrication offsets, gives $\hat r_t$.
Coverage calibration works as follows. On held-out data, we collect conformal scores, the standard quantity used to decide which labels enter a set, under simulated drift. We bin them by $\hat r_t$ and store one threshold per bin, taking the worst case over the bins adjacent to each one in the spirit of NAPS. At inference, the method reads $\hat r_t$, looks up a threshold, and emits a set. DOCTOR-style probing is triggered only when $\hat r_t$ exceeds the level at which the lookup's set size would pass a user budget.
The analytic part of the program tests the predicted residual scaling in $d$ and in the gradient. We will also derive the reference count $m$ needed for $\hat r_t$ to bound the true residual with a given probability.
Baselines
The naive reference point is uncorrected deployment of a model trained offline with noise injection. DOCTOR reports this collapsing from about 91% to about 57% under drift. Every method must at least recover this gap.
The strongest existing method is DOCTOR itself, which restores accuracy to above 90% at 0.14%–5% cycle overhead (DOCTOR). We will reimplement it in the same simulator, on the same networks and drift traces. Its trigger threshold and probe period will receive the same tuning budget as our bins and $m$. Beating it means matching its time-averaged accuracy within one point while invoking probing at least five times less often, under a matched full-system energy budget. A clear win would also be any drift trace on which it underperforms at equal overhead.
The second strong competitor is feedforward correction from on-chip temperature sensors. A per-tile thermistor reading is mapped through a fitted thermo-optic lookup table to a weight correction. For this sensor-feedforward approach, no reported result was found on Lacuna, so we will implement it ourselves with the same sensor density as reference rings.
SCATTER's phase-agnostic design serves as an architectural reference at its reported 91.31% on CNN-FMNIST, run under our dynamic traces.
Three ablations isolate the key ingredient. The first applies cancellation with no residual bound, using a fixed-threshold trigger. The second uses the residual-indexed lookup driven by the sensor reading instead of reference spread. The third uses a single global conformal threshold.
Why not just…?
Why not just read the on-chip temperature sensors and correct with a lookup table? This is the most serious threat. If drift is purely thermal and the sensors sit close to the rings, it could match reference cancellation at lower cost. Our argument is that a reference ring measures the optical quantity itself. It therefore absorbs laser-wavelength drift and the sensor-to-ring gradient, both of which a thermistor misses. That argument is unproven. The sensor-feedforward baseline and the sensor-driven ablation settle it. If they match ours in accuracy and coverage once laser drift is included in the simulation, we will report that the residual-budget idea survives but the reference ring does not.
Why not just lock everything with microheaters? Prior work documents that heater power can equal or exceed the optical compute power it protects (Kilocore ORNoC; GOA). We will include heater power in the energy comparison rather than dismiss it.
Why not just update a conformal threshold online as errors are observed? Online updating needs labels or error signals, which are unavailable during inference. It also reacts only after miscoverage has already happened. NAPS shows that unadjusted conformal sets fall to 40%–60% coverage under nuisance shift (NAPS). To make this comparison fair rather than a straw man, we will include an online-threshold variant that has oracle label access and measure its coverage specifically during hotspot onsets.
Proposed Experiments
Experiments use Fashion-MNIST (Fashion-MNIST) with the CNN used by SCATTER and DOCTOR, plus CIFAR-10 and CIFAR-100 with VGG-8 and ResNet-18. Each method runs over long drift traces. The metrics are time-averaged and worst-window accuracy, empirical coverage at 90% nominal and its worst window, average set size, probe invocations per hour, and full-system energy. Energy is computed with the converter and memory modeling of Architecture-level photonic modeling, including reference photodetectors and converters.
Three checks are decisive. First, the residual after cancellation must grow roughly linearly in $d$ and in the gradient. If it does not, the smooth-field premise is wrong and the method fails at its root. Second, the $\hat r_t$-indexed thresholds must hold coverage within two points of nominal in the worst hotspot window, where the global threshold falls well below. Third, against DOCTOR at matched energy, probing frequency must fall substantially with no accuracy loss. The adversarial sharp-gradient regime will tell us where cancellation stops paying for its rings, and we will report that crossover point.
References
- DOCTOR: Dynamic On-Chip Temporal Variation Remediation Toward Self-Corrected Photonic Tensor Accelerators
- Independent measurement of the neutrino mixing angle<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msub><mml:mrow><mml:mi>θ</mml:mi></mml:mrow><mml:mrow><mml:mn>13</mml:mn></mml:mrow></mml:msub></mml:mrow></mml:math>via neutron capture on hydrogen at Daya Bay
- SCATTER: Algorithm-Circuit Co-Sparse Photonic Accelerator with Thermal-Tolerant, Power-Efficient In-situ Light Redistribution
- An Efficient General-Purpose Optical Accelerator for Neural Networks
- Model-independent <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" altimg="si1.gif" overflow="scroll"><mml:msub><mml:mrow><mml:mover accent="true"><mml:mrow><mml:mi>ν</mml:mi></mml:mrow><mml:mrow><mml:mo stretchy="false">¯</mml:mo></mml:mrow></mml:mover></mml:mrow><mml:mrow><mml:mi>e</mml:mi></mml:mrow></mml:msub></mml:math> short-baseline oscillations from reactor spectral ratios
- Classification under Nuisance Parameters and Generalized Label Shift in Likelihood-Free Inference
- Architecture-Level Modeling of Photonic Deep Neural Network Accelerators
- An Optimized Low-Power Optical Memory Access Network for Kilocore Systems
- Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms