Adaptive Computation with Conformal Certification

1 INTRODUCTION

• Efficiency is also framed as a response to the energy, carbon, and capital costs of scaling, whose hardware returns are described as diminishing.The passage says algorithmic efficiency is critical for responsible scientific computing.Source: §1 · p. 1
• Conventional algorithmic optimizations can trade precision for runtime, a concern for sparse scientific tasks where errors may affect a small but consequential signal.The cited discussion contrasts these techniques with explicitly bounded risk on sparse signal.Source: §1 · pp. 1–2
• Sparse scientific signals such as tumors or rare astronomical transients can be missed even when aggregate metrics remain high, concentrating error costs in those regions.The authors argue that acceleration methods should explicitly bound risk on the sparse signal.Source: §1 · p. 2
• Conformal Adaptation selects candidate regions using conformal risk control, bounding the expected fraction of signal missed at a user-specified level while refining only selected regions.The calibrated bound is preserved through refinement under two stated conditions.Source: §1 · p. 2
• The framework combines conformalized candidate selection with sparse execution, admissible label aggregation, and case studies across scientific domains, data modalities, and sparsity regimes.The case studies quantify compute savings and empirically validate the guarantees.Source: §1 · p. 2

2 BACKGROUND

• Dense prediction maps spatially discretized inputs to outputs over a finite set of sites, with coarse partitions formed by unions of fine cells.The formulation covers pixels, voxels, and more generally indexed cells.Source: §2 · p. 2
• A hierarchical model decomposes prediction into a coarse model followed by refinement, which is described as cheaper than predicting directly at full resolution.The coarse and fine levels correspond to partitions of the domain.Source: §2 · p. 2
• Conformal risk control provides distribution-free, black-box prediction guarantees at coarse resolution under exchangeable calibration and test data and stated loss conditions.The loss is bounded, non-increasing, and right-continuous in its threshold.Source: §2 · p. 2
• Conformal Adaptation uses a coarse-resolution loss and hierarchical refinement, with CRC-based selection of candidate cells to reduce computation while maintaining fidelity.The passage identifies sparse selection as an additional efficiency step beyond lower-resolution prediction.Source: §2 · p. 3

3 CONFORMAL ADAPTATION

• The framework performs dense prediction at reduced spatial scale, then sparsely refines selected coarse cells using a CRC-calibrated threshold.The targeted operating regime has far fewer selected cells than coarse cells overall.Source: Figure 1 · p. 3, §3 · p. 3
• Refinement cost scales with the number of selected cells, while unselected cells receive null values or simple upsampling.Selection is derived from the coarse prediction and depends on test-time input information.Source: §3 · pp. 3–4
• An aggregation operator maps fine labels within each coarse cell to a non-negative signal weight and must assign zero to null-only cells while preserving positive weight for occupied cells.These conditions let cell-level weights represent signal for coarse-resolution calibration.Source: §3 · p. 4

3.1 THEORETICAL GUARANTEES

• Under exchangeability and admissible aggregation, CRC calibration bounds the expected fraction of true signal mass in unselected cells by α.The guarantee is established at coarse resolution using cell aggregates and the selected set.Source: §3 · p. 4, §3.1 · pp. 4–5
• The theorem states that dense and sparse pipelines agree within selected cells, while discarded-region signal mass is controlled and receives a marginal probability guarantee.The event that discarded cells contain no signal is exactly that all occupied cells were selected.Source: §3.1 · p. 5

4 PERFORMANCE METRICS

• The study evaluates dense classification and regression with DSC and lesion-detection metrics for segmentation, and MAE for spike-count and galaxy super-resolution prediction.It audits wall-clock speedup and FLOPs using both CPUs and GPUs.Source: §4 · p. 5
• Segmentation accuracy is assessed with Dice similarity coefficient and lesion-detection metrics.The passage identifies these as primary accuracy targets for segmentation tasks.Source: §4 · p. 5
• Mean absolute error is used for spike-count prediction and galaxy super-resolution.The evaluation also records timing on both CPUs and GPUs.Source: §4 · p. 5

5 CASE STUDY: CONFORMAL ADAPTATION FOR SEGMENTATION

• For brain-tumor masks, OR aggregation preserves positive tumor cells during coarsening and supports selection of regions for refinement.Each coarse cell is formed from fine sites; the OR operator is zero-preserving and monotonically increasing when any input is nonzero.Source: §5 · p. 5

5.1 BRAIN TUMOR SEGMENTATION

• The brain-tumor evaluation uses 484 cases, and the study applies Conformal Adaptation to U-Net, Swin-UNETR, and VISTA3D.Each architecture is trained at full and 4×-downsampled resolution, with the coarse model selecting sparse patches for a second-stage refiner.Source: §5.1 · p. 5
• 72× CPU wall-clock speedup and an 82.1× reduction in FLOPs are reported for VISTA3D brain-tumor inference.Figure 2 compares mean DSC and measured wall-clock speedups across devices.Source: Figure 2 · p. 6, §5.1 · p. 6

5.2 HEAD AND NECK CANCER AND METASTATIC LYMPH NODE SEGMENTATION

• Head-and-neck lesion detection has comparable recall and precision, but misses small adversarial GTVn cases.The predicted volume is used to detect lesions.Source: §5.2 · p. 7
• Once trained and calibrated, Coarse + Refine achieves DSC 0.72 versus 0.74 for Full while requiring 68× fewer FLOPs; test-time mirroring augmentation is used on all dense models.The figure shows comparable contours, with coarse inference reducing per-scan compute cost and sparse refinement improving prediction quality at little additional expense.Source: §5.2 · p. 7, Figure 5 · p. 7

6 CASE STUDY: CONFORMAL ADAPTATION FOR REGRESSION

• Regression aggregation sums non-negative fine-resolution targets within each coarse cell, so CRC bounds the expected fraction of signal mass in unselected regions.For regression scores outside [0, 1], a fixed strictly increasing map preserves score order and keeps selected sets nested in λ.Source: §6 · pp. 7–8

6.1 CASCADE: 1D CALCIUM-IMAGING SPIKE-RATE INFERENCE

• At α=0.1, 8.4% of coarse cells are retained with an analytic FLOPs speedup of 4.77×, while reconstructed-curve MAE stays within 0.05–0.10 Hz of baseline.Across α ∈ {0.01, 0.05, 0.1, 0.2}, test risk is ≤ α; median MAE is 0.250 Hz.Source: §6.1 · p. 8, Figure 6 · p. 8

6.2 GALAXY FLUX FIELD SUPER-RESOLUTION

• The galaxy pipeline uses an 8×-downsampled field and a sparse refiner whose normalized output is scaled by predicted median flux, making coarse-to-refined flux conservation exact.Source: §6.2 · p. 8
• 61.6× analytic FLOPs speedup over the full-resolution baseline is achieved while per-pixel MAE rises to 0.134 vs. 0.046 and total-flux relative error to 5.6% vs. 2.3%.At α=0.1, the pipeline retains 2.2% of coarse cells on average; Figure 7 shows measured wall-clock speedups across devices.Source: §6.2 · p. 8, Figure 7 · p. 9

• Static pruning, simplification, sparsification, and quantization evaluate error post hoc without accuracy guarantees, while their gains are largely orthogonal to input-adaptive sparsification.The passage gives examples including token sparsification for 3D segmentation and static resolution scaling.Source: §7 · p. 9
• Adaptive token selection, early exiting, and pixel refinement focus computation on informative regions but use learned or heuristic selection without guarantees on discarded information.The paper contrasts these approaches with its calibrated rule, which bounds discarded signal mass at risk level α.Source: §7 · p. 9
• Unlike conformal methods that use prediction sets around a fixed predictor, this approach uses the conformal certificate to guide computation.The passage also distinguishes prior adaptive-inference work that calibrates early exits to match full-model performance under compute constraints.Source: §7 · p. 9

Limitations

• The guarantee bounds expected signal mass in discarded cells, not refined-prediction error; it is marginal and can miss faint small lesions.It also assumes exchangeability between calibration and test data, making it sensitive to covariate shift across scanners or sites.Source: §8 · p. 9

Trade-offs

• Conformal Adaptation chooses where to compute, while quantization, pruning, sparse architectures, and token selection reduce the cost of each computation.Whether the savings compound is left as future work.Source: §8 · p. 9

Scope and future work

• The derivation extends to non-uniform domains such as graphs when refinement operators preserve the hierarchical partition.Source: §8 · p. 9

A.1 TRAINING

• Training proceeds in three steps: train the coarse model, extract sparse patches selected by an oracle or the trained coarse model, then train the refiner on those patches.Training at the calibrated threshold matches the refiner’s training distribution to inference, but requires one refiner per calibrated α.Source: §A.1 · p. 15

A.2 CALIBRATION

• CRC calibrates the coarse selection threshold using missed-signal-mass loss, which depends on coarse selection and cell weights, not on the refiner.Source: §A.2 · pp. 15–16

A.3 INFERENCE

• Inference downsamples the input, runs the coarse model, initializes a fine-resolution canvas from its output, then refines only selected coarse cells.Source: §A.3 · p. 16

B DERIVATION OF GUARANTEES

• The signal-mass loss lies in [0, 1], is non-increasing and right-continuous in λ, and is zero when every coarse cell is selected.Source: B DERIVATION OF GUARANTEES · p. 17
• Under exchangeability, the expected fraction of true signal mass in discarded cells is at most α, and the two pipelines agree at every fine site in selected cells.With probability at least 1−α/t, discarded cells carry at most a t-fraction of total signal mass for t in (α, 1].Source: B DERIVATION OF GUARANTEES · pp. 17–18
• The discarded signal mass is zero exactly when all occupied coarse cells are selected, for every admissible aggregation operator.Source: B DERIVATION OF GUARANTEES · pp. 17–18

C.1 MSD BRAIN TUMOR SEGMENTATION

• The MSD brain-tumor experiment uses a held-out 96-scan test subset and a separate 68-scan calibration set; tumor labels are preserved during 4 × 4 × 4 max-pooling.The scans contain four MRI sequences and voxel labels for three positive tumor subregions plus background.Source: §C.1.1 · p. 18, §C.1.2 · p. 19

C.1.5 QUALITATIVE EXAMPLES

• As α grows, Figure 8 shows the selected region shrinking toward the tumor boundary while measured cell miss rate rises toward its target.For two held-out scans, the Swin-UNETR seed-0 calibrated threshold is zero at α = 1% and 2%.Source: §C.1.5 · pp. 19–20, Figure 8 · p. 20

C.2.2 RESULTS

• The HNC Coarse + Refine results lose accuracy, especially for GTVn, with small lesions identified as a particular challenge.The authors describe the task as adversarial for cell-level miss control and suggest lesion-based CRC as future work.Source: §C.2.2 · p. 21

C.3 THRESHOLD SELECTION FOR SEGMENTATION

• A naïvely chosen threshold can yield high false-negative rates, and suitable thresholds vary across models and tasks, motivating held-out CRC calibration.Source: §C.3 · p. 21, Table 4 · p. 22

C.4 CASCADE CALCIUM SPIKE PREDICTION

• The CASCADE calcium-imaging experiment uses source-file-level train, calibration, and test splits, with CRC loss defined as missed spike-count mass.It uses 515 valid trials across 117 source files; coarse targets sum-pool raw spike counts by a factor of eight.Source: §C.4.1 · p. 23, §C.4.2 · p. 23

C.4.3 MODEL ARCHITECTURE

• The coarse 1D model matches the full-resolution baseline’s architecture and parameter count but runs on 128 rather than 1024 samples, reducing coarse-pass FLOPs by 8.0×.Source: §C.4.3 · p. 23
• The refiner predicts a non-negative spike count per retained block, then a mass-preserving Gaussian kernel reconstructs the continuous rate curve.This allows estimated count to spread into discarded neighboring blocks rather than hard-zeroing them in the displayed curve.Source: §C.4.3 · pp. 23–24
• The refiner’s total FLOPs depend on the number of selected windows, so Table 5 omits that total.Source: §C.4.3 · p. 24, Table 5 · p. 24

C.4.4 TRAINING AND CALIBRATION DETAILS

• At α = 0.1, CRC retains 8.4% of coarse blocks and the pipeline reports a 4.77× FLOPs speedup.Calibration uses the coarse model’s predicted τ=0.9 quantile score field; reconstructed-curve accuracy is evaluated by MAE against the fine-resolution target.Source: §C.4.4 · p. 24

C.4.5 RESULTS

• Tab. 6 reports reconstructed MAE in Hz against the 0.250 Hz full-resolution baseline, alongside retained fraction and final error across α values.The supplied passages do not state the table’s individual results.Source: §C.4.5 · p. 24, Table 6 · p. 24

C.5.1 DATASET DETAILS

• The galaxy-field study uses 3,000 simulated 1024×1024 fields, with 2,000/500/500 train, calibration, and test splits and object-disjoint catalog slices.Sources use analytic Sérsic fits from COSMOS rather than raw telescope pixels; this avoids amplifying correlated exposure noise when rescaling flux.Source: §C.5.1 · p. 25

C.5.2 DATA GENERATION PIPELINE

• Each source is rendered as a noisy, PSF-convolved input and a noise-free, sharply rendered target, making the task both deblurring and denoising.The coarse input is average-pooled, while the coarse target is sum-pooled so each cell represents total flux.Source: §C.5.2 · pp. 25–26

C.5.3 MODEL ARCHITECTURE

• The coarse model reuses the full baseline’s architecture and parameter count at 8× lower resolution, cutting its operations 64×; the refiner predicts within-cell flux distribution.The refiner’s normalized patch is multiplied by the coarse predicted total, conserving flux by construction; predicting networks use four monotone log-flux quantiles.Source: §C.5.3 · p. 26

C.5.4 TRAINING AND CALIBRATION DETAILS

• The refiner is trained on certified regions and their halo, using 716,629 cells from 2,000 fields, or 2.19% of all cells.Halo cells are included because they are near-empty and can show boundary-truncation errors.Source: §C.5.4 · pp. 26–27
• CRC scores cells by predicted 0.9-quantile flux, weights missed cells by flux mass, and selects the largest threshold whose corrected risk stays at or below α.A one-cell eight-connected halo is added before evaluating loss.Source: §C.5.4 · p. 27
• At α=0.1, test risk is 0.040 (0.042 corrected) with 2.18% of cells certified, yielding 61.6× fewer floating-point operations than the full-resolution baseline.This is close to the 64× ceiling from the coarse pass; refinement costs about 4% of the coarse-pass cost.Source: §C.5.4 · p. 27
• At the same α, the calibrated coarse model certifies 1.53× as many cells as an oracle using true coarse flux totals.The oracle certifies 1.43% of cells at risk α under the same halo.Source: §C.5.4 · p. 27

C.5.5 RESULTS

• Figure 10 presents a selection of galaxy flux field predictions.The supplied passages do not describe a specific comparison or outcome in the figure.Source: §C.5.5 · p. 27, Figure 10 · p. 28