Conformally Calibrated Adaptive Computation in Neural Operators
Organizations: Deutsches Elektronen-Synchrotron DESY, Hamburg, Germany · Department of Information Engineering, Computer Science and Mathematics University of L’Aquila, L’Aquila, Italy
Abstract
Selecting less computation in a neural operator need not reduce inference time. We study this gap using conformal calibration to choose Fourier modes or channel width per input subject to a fixed error budget. A triangle-inequality envelope bounds all candidates simultaneously; a direct-score rule handles non-monotone candidate errors and, with an input-independent score, calibrates a static baseline under the same marginal guarantee. Under frozen-model and exchangeability assumptions, both rules bound the probability of issuing a prediction outside the budget and abstain when calibration is infeasible. On Burgers and Darcy flow across three training seeds, mode selection reduces active spectral support without a measurable latency reduction, because full-grid transforms and channel mixing dominate the cost. Narrowing channels does reduce Darcy forward cost, but reference-based selection must first compute the dense prediction. An input-only binary width selector avoids that pass and runs at 0.52-0.95 times dense latency; yet static widths calibrated under the same rule are often as fast or faster, and on a separately generated 1,000-example test set they generally achieve higher coverage, while the selector falls below its 90% target in more than half of the settings. Adaptive computation in neural operators should therefore be judged by achieved coverage and complete runtime against a matched static comparator.
Figures & tables
| Approach | Statistical role | Execution finding |
|---|---|---|
| Reference-based spectral | Simultaneous error envelope for mode selection | Little or no latency change in the tested Burgers and Darcy sweeps (Part I) |
| Reference-based width | Direct-score selection over nested widths | Requires the dense pass, so no latency saving (Part II) |
| Binary TV selector | Input-only binary gate: widths , scored by | Full pipeline: – dense latency on batched Darcy (Part II) |
| Static widths | Calibrated within each model family over | Runtime advantage depends on model family and tolerance (Part II) |
| Track | Model selection / development | Calibration | Evaluation | Status |
|---|---|---|---|---|
| Spectral, Burgers | Validation ( ) | Test half ( ) | Other test half ( ) | Held out from model selection |
| Spectral, Darcy | Validation ( ) | Fresh canonical pool ( ) | New Darcy set ( ) | Held out from model selection |
| Early width analyses | Validation and original test | Validation or fresh pool ( ) | Original test ( ) | Development |
| Main width coverage | Policies locked before generation | Fresh canonical pool ( ) | New Darcy set ( ) | New examples; prior development reuse |
| Runtime | Locked width policies | Same fitted thresholds | Original canonical batches | Hardware benchmark |
| Symbol | Meaning |
|---|---|
| Input field and target solution | |
| Candidate and full spectral support; candidate and full width | |
| Predictions at the selected support or width | |
| Absolute prediction errors; is the full-reference error | |
| Absolute disagreement with the corresponding full prediction | |
| Fixed absolute error budget and allowed marginal error-event probability |
| Batch 1 | Batch 20 | |
|---|---|---|
| 2 | 1.816 0.022 | 11.032 0.029 |
| 4 | 1.808 0.022 | 11.045 0.047 |
| 6 | 1.838 0.262 | 11.054 0.042 |
| 8 | 1.812 0.021 | 11.066 0.045 |
| 10 | 1.808 0.027 | 11.097 0.045 |
| 12 | 1.821 0.023 | 11.114 0.066 |
| Triangle rule | Direct-score rule | ||||
|---|---|---|---|---|---|
| Coverage | Support fraction (%) | Coverage (%) | Support fraction (%) | ||
| 0.02 | 0.10 | 99.0–100.0% | 52.2 / 54.2 / 55.0 | 92.0 / 88.0 / 94.0 | 46.0 / 45.1 / 49.8 |
| 0.03 | 0.10 | 100.0% | 44.2 / 44.9 / 47.2 | 95.0 / 94.0 / 91.0 | 41.2 / 41.5 / 43.2 |
| 0.05 | 0.10 | 100.0% | 39.2 / 39.4 / 40.8 | 83.0 / 85.0 / 83.0 | 35.6 / 36.8 / 38.1 |
| 0.10 | 0.10 | 100.0% | 30.6 / 31.4 / 32.0 | 89.0 / 89.0 / 88.0 | 28.9 / 29.1 / 29.8 |
| 0.03 | 0.05 | 100.0% | 45.2 / 45.5 / 48.0 | 97.0 / 99.0 / 95.0 | 41.5 / 42.1 / 43.8 |
| Triangle rule | Direct-score rule | |||
|---|---|---|---|---|
| Coverage | Support fraction (%) | Coverage (%) | Support fraction (%) | |
| 0.05 | 99.6% | 99.7 / 99.9 / 99.9 | 85.4 / 88.0 / 86.7 | 94.9 / 97.0 / 96.2 |
| 0.08 | 99.7–100.0% | 93.0 / 97.4 / 95.4 | 89.1 / 90.1 / 87.7 | 84.5 / 90.5 / 86.9 |
| 0.12 | 99.6–99.7% | 71.1 / 81.1 / 76.0 | 90.8 / 93.1 / 90.2 | 59.8 / 69.7 / 64.0 |
| 0.18 | 99.5–99.8% | 37.6 / 45.0 / 40.1 | 92.0 / 91.9 / 88.4 | 30.2 / 35.8 / 31.3 |
| Seed | Nested routing coverage / w8 / time | Nested static width/coverage/time | Independent routing coverage / w8 / time | Independent static width/coverage/time | |
|---|---|---|---|---|---|
| 42 | 0.020 | 89.6% / 34.6% / 0.884 | 24 / 92.4% / 0.792 | 89.9% / 43.6% / 0.786 | 24 / 93.2% / 0.792 |
| 42 | 0.024 | 87.3% / 73.5% / 0.708 | 16 / 95.6% / 0.586 | 89.2% / 83.9% / 0.589 | 16 / 96.6% / 0.586 |
| 42 | 0.028 | 90.5% / 94.2% / 0.536 | 16 / 97.3% / 0.586 | 91.5% / 100.0% / 0.466 | 8 / 91.5% / 0.384 |
| 43 | 0.020 | infeas. | infeas. | 89.7% / 38.3% / 0.834 | 24 / 95.2% / 0.792 |
| 43 | 0.024 | 89.3% / 61.8% / 0.775 | 16 / 91.6% / 0.586 | 89.5% / 73.7% / 0.686 | 16 / 93.9% / 0.586 |
| 43 | 0.028 | 87.6% / 96.3% / 0.520 | 16 / 95.7% / 0.586 | 90.7% / 98.5% / 0.479 | 16 / 96.4% / 0.586 |
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
| Tolerance | Mean | Median | At | Support fraction |
|---|---|---|---|---|
| relative-slack 0.10 | 14.91 | 16 | 57.0% | 93.2% |
| relative-slack 0.25 | 13.53 | 14 | 17.5% | 84.6% |
| relative-slack 0.50 | 12.33 | 12 | 3.0% | 77.1% |
| fixed | 10.53 | 10 | 9.5% | 65.8% |
| fixed | 12.61 | 12 | 25.0% | 78.8% |
| Dataset | mean | vs. in-distribution | Mean | Held-out coverage | Target |
|---|---|---|---|---|---|
| In-distribution, | 0.0042 | 7.31 | 100.0% | 90% | |
| Out-of-distribution, | 1.074 | 7.31 | 0.0% | 90% | |
| Out-of-distribution, | 4.865 | 11.90 | 0.0% | 90% |
| Nested checkpoint | Independent models | |||||
| 42 | 43 | 44 | 42 | 43 | 44 | |
| 0.020 | 24 | – | 24 | 16 | 24 | 24 |
| 0.024 | 16 | 16 | 16 | 16 | 16 | 8 |
| 0.028 | 16 | 16 | 16 | 8 | 16 | 8 |
| Nested checkpoint | Independently trained | |||||||
|---|---|---|---|---|---|---|---|---|
| Seed | ||||||||
| 42 | 0.021301 | 0.014151 | 0.012612 | 0.012098 | 0.019331 | 0.013280 | 0.012634 | 0.012107 |
| 43 | 0.021760 | 0.016278 | 0.014320 | 0.013673 | 0.020660 | 0.014789 | 0.011494 | 0.012116 |
| 44 | 0.022569 | 0.013855 | 0.012242 | 0.011429 | 0.015863 | 0.015779 | 0.013298 | 0.012040 |
| Seed | Mean chosen error / dense error | Frac. at width 8 | |
|---|---|---|---|
| 42 | 0.020 | ||
| 42 | 0.024 | ||
| 42 | 0.028 | ||
| 43 | 0.020 | infeasible | – |
| 43 | 0.024 | ||
| 43 | 0.028 |
| Static coverage | Selector coverage | Selector cost proxy | |
|---|---|---|---|
| 0.020 | – | – | – |
| 0.024 | – | – | – |
| 0.028 | – | – | – |
| Seed | Static width | Static coverage | Static / dense | Selector / dense | |
|---|---|---|---|---|---|
| 42 | 0.020 | 16 | 93.5% | 0.586 | 0.786 |
| 42 | 0.024 | 16 | 96.6% | 0.586 | 0.589 |
| 42 | 0.028 | 8 | 91.5% | 0.384 | 0.466 |
| 43 | 0.020 | 24 | 95.2% | 0.792 | 0.834 |
| 43 | 0.024 | 16 | 93.9% | 0.586 | 0.686 |
| 43 | 0.028 | 16 | 96.4% | 0.586 | 0.479 |
| Seed | Batch | Dense (ms) | Selector (ms) | Local frac. @ w8 | Proxy ratio / Measured ratio | |
|---|---|---|---|---|---|---|
| 42 | 0.020 | 20 | 11.08 | 9.29 | 45.0% | 0.730 / 0.838 |
| 42 | 0.020 | 64 | 34.59 | 27.05 | 37.5% | 0.768 / 0.782 |
| 42 | 0.020 | 128 | 64.72 | 52.86 | 39.1% | 0.761 / 0.817 |
| 42 | 0.024 | 20 | 11.10 | 8.44 | 55.0% | 0.670 / 0.761 |
| 42 | 0.024 | 64 | 34.70 | 21.10 | 65.6% | 0.595 / 0.608 |
| 42 | 0.024 | 128 | 64.77 | 38.52 | 72.7% | 0.555 / 0.595 |
Adaptive Computation
Adaptive computation allocates model depth, tokens, attention, or sensing effort according to input difficulty and resource constraints.
See all 117 papers