Pith. sign in

REVIEW 2 major objections 6 minor 40 references

This paper claims that most known tricks for speeding up large FIR filters compose into one searchable design language, and that automatically searching it yields approximations that are both faster and more accurate than hand-tuned methods

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 09:47 UTC pith:YFWL53FB

load-bearing objection A genuinely useful synthesis-and-search system for fast filter approximations, with a real headline-number inconsistency the authors need to fix before it can be taken at face value. the 2 major comments →

arxiv 2607.20634 v1 pith:YFWL53FB submitted 2026-07-22 cs.GR

Finding Fast Filters

Karima Ma , Andrew Adams , Jonathan Ragan-Kelley This is my paper
classification cs.GR
keywords filter approximationprogram synthesisPareto frontiertail-canceling IIRmulti-rate filteringgradient descentimage processingaudio processing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper's central claim is that the familiar cost-saving techniques for large FIR filters — multi-rate down/upsampling, recursive IIR and tail-canceling TIIR filters, frequency-response masking with strided FIRs, and composition by cascade and sum — are all instances of one design language. Given a target filter, the system enumerates tree-structured programs in this language, fits their continuous parameters with gradient descent, predicts throughput with a cost model, and compiles the top designs to optimized vectorized parallel C++. Against Gaussians, Lanczos, head-related impulse responses, telephone effects, Gabors, and low-pass filters, the search produces approximations that match exact filtering while running tens to hundreds of times faster than FFT convolution, and it Pareto-dominates the prior methods (convolution pyramids, frequency-response masking, triple-box blur, YVV, rank-2 SVD) when those are implemented with the same compiler. If correct, this turns filter acceleration from hand-derived designs into automated search over a structured program space that spans imaging and audio.

Core claim

In the paper's own terms, the discovery is that fast filter approximation reduces to a differentiable program search: every well-known acceleration scheme is a composition of a small set of primitives — dense FIRs, tail-canceling IIRs (recursive filters whose infinite tail is canceled by a sparse correction, yielding a finite impulse response at near-constant cost), strided frequency-response-masking blocks, down/upsample wrappers, cascades, sums, and a pipe operator that promotes 1D filters to 2D. The system enumerates legal program trees under hand-written constraints informed by signal-processing intuition, fits each structure's continuous parameters by gradient descent (using a time-step

What carries the argument

The load-bearing object is a small domain-specific language whose primitives are composable filter programs: FIR (dense finite impulse response), TIIR (tail-canceling IIR — a recursive filter plus a sparse FIR correction that truncates an infinitely long impulse response after N steps at O(1) cost per output), Stride (a strided child computation followed by a factored-tower interpolator, generalizing frequency-response masking), DownUpsampleI/II (prefilter, subsample, child, zero-upsample, interpolate), and binary combinators Cascade, Sum, and Pipe that build 2D filters from 1D ones. The subtle piece is the differentiable TIIR parameterization: rather than differentiating through the recurre

Load-bearing premise

The reported Pareto frontier is only as trustworthy as the hand-written enumeration rules plus the cost model, which ranked 150k sampled designs and benchmarked only 2.5k; if the cost model mispredicts a pruned structure's throughput, the true frontier may contain better designs the search never saw.

What would settle it

Compile and benchmark all ~150k sampled designs for one target (e.g., 2D Gaussian size 201) instead of only the top-3 Pareto tiers. If any unbenchmarked design beats every published frontier point in throughput at equal or higher PSNR, the cost-model pruning is the limiting factor and the reported frontier is incomplete. A second test: run the identical pipeline on a held-out filter family not used in the paper, such as a sharp band-pass with 60 dB stopband attenuation, and check whether any generated design exceeds FFT throughput at PSNR above 50 dB; failure would mean the search's breadth do

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • For large Gaussian blurs, the search finds approximations that are visually indistinguishable from ground truth while running up to about 600x faster than FFT convolution on cache-resident images.
  • The commonly used triple-box Gaussian approximation is Pareto-dominated: generated designs are 3.9–4.7x faster at equal or higher PSNR on 2D Gaussians.
  • For real-time audio, HRIR binaural processing is reported 2.37x faster than a 16k-sample FFT and 8.66x faster than a 256-sample FFT block, with output indistinguishable from exact filtering; a telephone effect filter is 2.74x faster than FFT.
  • For 2D low-pass filters that are not separable, generated approximations reach speedups up to about 613x over FFT for size 401.
  • Prior acceleration methods appear as special cases or points inside the same search space, so isolated hand designs for new filters are likely to be automatically rediscovered or beaten by search.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The one-time training cost per filter could be amortized if, as the paper hints, a trained design can be resized across a parameter family (e.g., Gaussians of different radii); that would make the approach practical for production filter libraries.
  • The reported dominance suggests that classical operation-count comparisons understate real gaps: on modern CPUs, vectorization, fusion, and memory locality matter as much as arithmetic count, a lesson likely to transfer to other signal-processing kernels.
  • Because the enumeration constraints are hand-written, the breadth of the method is only as general as those rules; a strong test is whether the same constraints—or a learned variant—transfer to targets far outside the tested set, such as very sharp band-pass filters or 3D convolution kernels.
  • The cost model's 0.97–0.98 correlation bounds how much of the true design space the frontier explores; benchmarking all sampled designs rather than the top 2,500 would reveal how many potentially better programs were pruned.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. The paper presents a DSL for fast FIR approximations whose primitives unify multi-rate filtering, tail-canceling IIRs, cascades/sums, and stride/interpolation. Search enumerates program structures, fits continuous parameters with gradient descent (including a reparameterization for TIIRs), uses a cost model to rank designs, then lowers top designs to optimized C++ (SIMD, tiling, fusion). Evaluation covers 1D Gaussian, Lanczos, HRIR, and telephone filters, and 2D Gaussian, Gabor, and low-pass filters, reporting Pareto frontiers against CP, FRM, triple-box, YVV, and SVD baselines, plus exact FFT/direct/separable methods.

Significance. If the results hold, this is a substantial systems contribution: the DSL subsumes CP and FRM, the differentiable TIIR parameterization is a useful technical building block, and the automatic lowering to fast C++ addresses a real gap in prior approximate-filtering work. The evaluation is extensive, includes qualitative outputs, and is often conservative (e.g., overestimating YVV throughput for Gabors). The central search-and-lowering pipeline is coherent, and the claimed dominance over the investigated baselines is credible. However, the headline 600x speedup in Fig. 1 is not supported as stated, and the completeness of the reported Pareto frontier is conditional on cost-model rankings that are not independently validated.

major comments (2)
  1. [Fig. 1; §4.6; Fig. 38] The headline claim that the Gaussian approximation runs '600 times faster than the fastest exact filtering method' is internally inconsistent. Section 4.6 explicitly states that for 2D Gaussians 'linearly separable convolution is faster' than FFT, and Fig. 38 reports a maximum speedup of 68.6x over that separable baseline for size 401. Thus 600x is measured against FFT convolution, not the fastest exact method. This is not a cosmetic caption issue: it inflates the advertised speedup by roughly 8.5x. Please correct the caption/text to say '600x faster than FFT convolution' or, preferably, report speedups against the actual fastest exact baseline (separable convolution), which yields ~70x.
  2. [§3.3; §7.2] The paper's 'Pareto frontier' is computed after pruning by enumeration constraints (Appendix 7.2) and by a cost model (Section 3.3) that ranks 150k designs and keeps only the top 3 tiers (2,500 designs) for benchmarking. The cost model has 0.97–0.98 log-log correlation, but no validation is provided that the cost model or the enumeration rules do not discard designs that would actually lie on the true frontier. Since the abstract and Section 3 claim to 'generate a Pareto frontier,' the paper should soften this wording to 'frontier over sampled/ranked designs' or add a validation experiment (e.g., benchmarking a random sample of lower-tier and pruned designs) to demonstrate that no Pareto-optimal designs are systematically excluded.
minor comments (6)
  1. [Fig. 1 caption] Typo: 'one our programs' should be 'one of our programs.' Also, the caption should specify which exact method the 600x figure refers to.
  2. [§3.3 vs Fig. 17] The text reports correlation coefficients 0.98 (1D) and 0.97 (2D), but Fig. 17 labels them as r=0.97 (1D) and r=0.98 (2D). The numbers are swapped and should be corrected.
  3. [§4.8] Typo: 'PSNR 61.dB' should be 'PSNR 61.4dB' or similar.
  4. [Table 4] The first data row 'Ours 1 1.96 21.94 46.97' is misaligned and confusing; align the columns and clarify the row labels.
  5. [§7.4] Capitalization of 'DownUpSample' vs. 'DownUpsample' is inconsistent throughout the manuscript and appendix; unify the notation.
  6. [General] The paper would benefit from an artifact/reproducibility statement, including whether code and benchmark scripts will be released.

Circularity Check

0 steps flagged

No circular derivation: search outputs are benchmarked, not predicted; the DSL subsuming CP/FRM is an expressiveness claim, not a forced result; only non-load-bearing self-citation exists.

full rationale

The central claim—that search finds faster, higher-quality approximations than CP/FRM/triple-box/YVV—is supported by direct benchmarking, not by the fitted cost model alone. Section 3.3 states that the cost model is used only to prune the search ('Afterwards, a small subset, 2500 out of 150k total designs, that made it to the top 3 Pareto tiers for each target were compiled and benchmarked to produce the final Pareto frontier'). The final Pareto points are therefore measured, not predicted from fitted constants, so the cost-model fits are not being relabeled as predictions. Similarly, the fact that CP and FRM are subsumed in the DSL by construction (Section 4: 'Both CP and FRM exist in our search space') explains why the system can express those baselines, but it does not by itself make the search results superior; the dominance is an empirical result. The self-citation to Ma et al. 2022 appears in related-work context ('as was done in Ma et al. [2022]') and in a discussion of Halide's limitations; it is not the load-bearing justification for the paper's originality or correctness. The TIIR derivation and the truncation reparameterization are developed in the paper with equations, not imported from a self-citation. One non-circular inconsistency should be noted: Fig. 1 and the introduction advertise '600 times faster than the fastest exact filtering method' for a 2D Gaussian, while Section 4.6 states that linearly separable convolution is faster than FFT for 2D Gaussians and reports the largest separable-relative speedup as 70.2x. This undermines the headline quantitative claim, but it is a measurement/comparison error, not a circular derivation. Overall, no step of the claimed derivation reduces to its own inputs or to a self-citation chain.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The central claims rest on several hand-chosen hyperparameters (γ, pole priors, annealing schedule, cost-model constants, early stopping) and domain assumptions about TIIR optimization, enumeration pruning, and cost-model fidelity. No new physical entities are introduced.

free parameters (6)
  • TIIR time-step resolution γ = ≥ max target frequency, exact per-target values not listed
    Introduced in §3.4.2 to rescale pole magnitudes into the unit circle; hand-set per target and central to TIIR training stability.
  • Pole-type sampling probabilities = [0.875, 0.1, 0.025]
    §3.5.3: hand-chosen prior over complex-conjugate / distinct-real / duplicate-real poles; influences which TIIRs are discovered.
  • Sigmoid temperature annealing τ = 5 → 50, rate 1.002 per iteration
    §3.5.3: controls soft truncation windows during TIIR training.
  • Cost-model FIR/TIIR coefficients = FIR: 0.5/0.8/0.9; TIIR table [24, 19, 16, 13, 9]
    §7.3: constants determined from benchmarking on the target hardware and used to select which designs are compiled/benchmarked for the final Pareto frontier.
  • Early-stopping PSNR threshold and minimum iterations = not specified numerically
    §3.5.2: hand-set termination criteria that determine which designs survive training and enter the final frontier.
  • Training phase tile size D = D = W/2, rounded to a multiple of S
    §3.5.1: chosen for training efficiency; affects the phase coverage and loss estimate.
axioms (6)
  • standard math Z-transform and partial-fraction decomposition for order-2 IIRs (Eqs. 11-12)
    Used in §3.4.2 to express TIIR impulse responses as weighted sums of cosine/sine or exponentials.
  • standard math TIIR algebra of Eqs. (5)-(10) from Wang and Smith 1997
    Basis for representing truncated FIRs as sparse FIR + low-order IIR; central to the TIIR primitive.
  • domain assumption Useful TIIRs have poles close to the unit circle and γ-rescaling makes the unit circle filled with useful magnitudes
    §3.4.2 / Figs. 3-4: the training stabilization method relies on this assertion; no formal guarantee is provided.
  • domain assumption Enumeration constraints in Appendix 7.2 eliminate only invalid or unlikely-competitive structures
    Used throughout the search to prune the space; if a constraint is wrong, good designs could be missed.
  • domain assumption Cost model ranks designs accurately enough that top-3 Pareto tiers contain the true Pareto frontier
    §3.3: only 2,500 of 150k designs are benchmarked based on cost-model ranking; correlation 0.97-0.98 is on sampled designs, not a proof of no missed designs.
  • domain assumption Training on all-phase impulse inputs generalizes to white-noise PSNR
    §3.5.1: loss is computed on specially constructed phase impulses; quality is later measured on white noise, so transfer is assumed.

pith-pipeline@v1.3.0-alltime-deepseek · 54691 in / 14075 out tokens · 123084 ms · 2026-08-01T09:47:06.945055+00:00 · methodology

0 comments
read the original abstract

Processing images, video, and audio often requires running large finite impulse response (FIR) filters with strict performance and latency requirements. Prior methods for fast filter approximations are special cases or combinations of a few key techniques: multi-rate and recurrent filtering, and decomposing filters into sums or cascades. We unify these techniques as primitives within a single design language for fast 1D and 2D filters. Given a target filter to approximate, we automatically search this program space, fitting continuous parameters with gradient descent, to generate a Pareto frontier of algorithms that trade off performance with quality. Our system produces substantially higher-quality and faster filter approximations than have been previously described for several popular imaging and audio filters. Furthermore we demonstrate how to automatically lower programs in this design space to optimized, vectorized, parallel, C++ code which is fused for data locality.

Figures

Figures reproduced from arXiv: 2607.20634 by Andrew Adams, Jonathan Ragan-Kelley, Karima Ma.

Figure 1
Figure 1. Figure 1: We unify fast filter approximation techniques as primitives within a single design language. This allows us to search over programs to find novel [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: TIIRs extract FIRs from IIRs. Left: an exponentially decaying first [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The time step resolution parameter s rescales any desired range of [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The most useful TIIRs use IIRs that decay slowly. In pole-space (top [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Truncation regions for TIIRs are learned as the inflection locations of increasing and decreasing sigmoids that mask values in the zeros regions. During training we gradually increase the sigmoid temperatures, steepening the transition regions, to encourage integer cutoff locations. From top to bottom: original TIIR ℎ[𝑛], the sigmoids, and the truncated TIIR ℎ[𝑛] ′ . through modulo to perform truncation by… view at source ↗
Figure 6
Figure 6. Figure 6: Convolution by a truncated infinite impulse response of size [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Designs that use multi-rate filtering are not shift-invariant. Therefore, [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Our search space subsumes FRM and CP. FRM designs are a single [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Throughput vs. quality for approximations of 1D Gaussian sizes 101, 201, 401, and 801. We show the Pareto frontier across all methods (designs on the [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Throughput vs. quality for approximations of 1D Lanczos filter sizes 101, 201, 401, and 801. We show the Pareto frontier across all methods (designs on [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Measurement setup for the HRIR dataset. The subject is seated in [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Throughput vs. quality for approximations of each HRIR section for the right (top row) and left (bottom row) ear. We show the Pareto frontier across [PITH_FULL_IMAGE:figures/full_fig_p016_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Throughput vs. quality for approximations of each 1D Telephone filter section. We show the Pareto frontier across all methods (designs on the blue [PITH_FULL_IMAGE:figures/full_fig_p016_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Throughput vs. quality for 2D Gaussian approximations for sizes 101, 201, and 401. We show the Pareto frontier across all methods (designs on the blue [PITH_FULL_IMAGE:figures/full_fig_p018_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Throughput vs. quality for 2D Gabor filter approximations for sizes 101, 201, and 401. We show the Pareto frontier across all methods (designs on the [PITH_FULL_IMAGE:figures/full_fig_p019_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Throughput vs. quality for 2D Lowpass approximations for sizes 101, 201, and 401. We show the Pareto frontier across all methods (designs on the blue [PITH_FULL_IMAGE:figures/full_fig_p020_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Given a design, our cost model predicts machine cycles spent per [PITH_FULL_IMAGE:figures/full_fig_p023_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Figures 18-21 show the responses for selected models across all 1D Gaussian sizes. For each size we show a lower and higher quality Pareto frontier [PITH_FULL_IMAGE:figures/full_fig_p027_18.png] view at source ↗
Figure 22
Figure 22. Figure 22: Figures 21-24 are the response plots for 1D Lanczos filters sizes 101, 201, 401, and 801. The takeaways are similar to those for 1D Gaussians. FRM [PITH_FULL_IMAGE:figures/full_fig_p029_22.png] view at source ↗
Figure 26
Figure 26. Figure 26: Figures 25-27 show the responses for each section of the right ear’s HRIR. Figures 28-30 show the responses for the left ear’s sections. The -50dB line is [PITH_FULL_IMAGE:figures/full_fig_p031_26.png] view at source ↗
Figure 32
Figure 32. Figure 32: Figures 31-34 show the responses for each section of the Telephone filter. We show the group delay instead of the phase response because this is [PITH_FULL_IMAGE:figures/full_fig_p034_32.png] view at source ↗
Figure 36
Figure 36. Figure 36: Figures 35-37 show the responses for 2D Gaussian sizes 101, 201, and 401. For all sizes, triple box blur has ringing and non-circular impulse response. It [PITH_FULL_IMAGE:figures/full_fig_p037_36.png] view at source ↗
Figure 39
Figure 39. Figure 39: Figures 38-40 are the response plots for 2D Gabor sizes 101, 201, and 401. CP struggles to get perfect approximations up to -50dB because Gabor filters [PITH_FULL_IMAGE:figures/full_fig_p040_39.png] view at source ↗
Figure 42
Figure 42. Figure 42: Figures 41-43 show the response plots for 2D Low-pass filters sizes 101, 201, and 401. For all sizes both of our designs have comparble or higher [PITH_FULL_IMAGE:figures/full_fig_p042_42.png] view at source ↗
Figure 45
Figure 45. Figure 45: Figures 45 - 47 show outputs for 2D size 101 Gaussian filter approximations, on patches from different images taken with the authors’ iPhone 13 Pro. [PITH_FULL_IMAGE:figures/full_fig_p044_45.png] view at source ↗
Figure 46
Figure 46. Figure 46: 2D Gaussian filter, size 101 approximation outputs. [PITH_FULL_IMAGE:figures/full_fig_p045_46.png] view at source ↗
Figure 47
Figure 47. Figure 47: 2D Gaussian filter, size 101 approximation outputs. [PITH_FULL_IMAGE:figures/full_fig_p046_47.png] view at source ↗
Figure 48
Figure 48. Figure 48: Figures 48 and 49 show outputs for 2D size 201 Gaussian filter approximations on patches from different images taken with the authors’ iPhone 13 Pro. [PITH_FULL_IMAGE:figures/full_fig_p047_48.png] view at source ↗
Figure 49
Figure 49. Figure 49: 2D size 201 Gaussian filter approximation outputs. [PITH_FULL_IMAGE:figures/full_fig_p048_49.png] view at source ↗
Figure 50
Figure 50. Figure 50: Outputs for 2D size 401 Gaussian filter approximations on a patch from an image taken with the authors’ iPhone 13 Pro. For such large blur sizes the [PITH_FULL_IMAGE:figures/full_fig_p049_50.png] view at source ↗
Figure 51
Figure 51. Figure 51: Figures 51 - 53 show outputs for 2D size 101 Lowpass filter approximations, on patches from image: [PITH_FULL_IMAGE:figures/full_fig_p050_51.png] view at source ↗
Figure 52
Figure 52. Figure 52: 2D size 101 Lowpass filter approximation outputs. [PITH_FULL_IMAGE:figures/full_fig_p051_52.png] view at source ↗
Figure 53
Figure 53. Figure 53: 2D size 101 Lowpass approximation outputs. [PITH_FULL_IMAGE:figures/full_fig_p052_53.png] view at source ↗
Figure 54
Figure 54. Figure 54: Figures 55 and 56 show outputs for 2D size 201 Lowpass filter approximations, on patches from different images taken with the authors’ iPhone 13 [PITH_FULL_IMAGE:figures/full_fig_p053_54.png] view at source ↗
Figure 55
Figure 55. Figure 55: 2D size 201 Lowpass filter approximation outputs. [PITH_FULL_IMAGE:figures/full_fig_p054_55.png] view at source ↗
Figure 56
Figure 56. Figure 56: Outputs for 2D size 401 Lowpass filter approximations on patches from an image taken with the authors’ iPhone 13 Pro. Our model produces outputs [PITH_FULL_IMAGE:figures/full_fig_p055_56.png] view at source ↗
Figure 57
Figure 57. Figure 57: 2D size 101 Gabor filter approximation outputs on an image patch taken from [PITH_FULL_IMAGE:figures/full_fig_p056_57.png] view at source ↗
Figure 58
Figure 58. Figure 58: 2D size 201 Gabor filter approximation outputs on image patches from a photo taken with the authors’ iPhone 13 Pro. Outputs are visually [PITH_FULL_IMAGE:figures/full_fig_p057_58.png] view at source ↗
Figure 59
Figure 59. Figure 59: Figures 59 - 61 show 2D size 401 Gabor filter approximation outputs on patches from photos taken by the authors’ iPhone 13 Pro. CP outputs have [PITH_FULL_IMAGE:figures/full_fig_p058_59.png] view at source ↗
Figure 60
Figure 60. Figure 60: 2D size 401 Gabor filter approximation outputs. [PITH_FULL_IMAGE:figures/full_fig_p059_60.png] view at source ↗
Figure 61
Figure 61. Figure 61: 2D size 401 Gabor filter approximation outputs. [PITH_FULL_IMAGE:figures/full_fig_p060_61.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 8 canonical work pages

  1. [5]

    InIEEE International Conference on Acoustics, Speech and Signal Process- ing, ICASSP 2022, Virtual and Singapore, 23-27 May

    Direct Design of Biquad Filter Cascades with Deep Learning by Sampling Random Polynomials. InIEEE International Conference on Acoustics, Speech and Signal Process- ing, ICASSP 2022, Virtual and Singapore, 23-27 May

    2022

  2. [6]

    https://doi.org/10.1109/ICASSP43922.2022.9747660 R

    IEEE, Singapore, 3104–3108. https://doi.org/10.1109/ICASSP43922.2022.9747660 R. Crochiere and L. Rabiner

    arXiv 2022

  3. [7]

    https: //doi.org/10.1109/ICASSP.1982.1171853 Elhanan Elboher and Michael Werman

    ICASSP, Paris, 2046–2049. https: //doi.org/10.1109/ICASSP.1982.1171853 Elhanan Elboher and Michael Werman

    arXiv 2046

  4. [13]

    https://doi.org/10.1145/2813885.2737977 Jonathan Frankle and Michael Carbin

    Synthesizing data structure transformations from input-output examples.SIGPLAN Not.50, 6 (June 2015), 229–239. https://doi.org/10.1145/2813885.2737977 Jonathan Frankle and Michael Carbin

    arXiv 2015

  5. [20]

    Graph.41, 5, Article 172 (May 2022), 18 pages

    Searching for Fast Demosaicking Al- gorithms.ACM Trans. Graph.41, 5, Article 172 (May 2022), 18 pages. https: //doi.org/10.1145/3508461 Henry Massalin

    work page doi:10.1145/3508461 2022

  6. [23]

    https://doi.org/10.1016/S0146-664X(81)80009-3 Diego Nehab, André Maximo, Rodolfo S

    Box-filtering techniques.Computer Graphics and Image Processing 17, 1 (1981), 65–70. https://doi.org/10.1016/S0146-664X(81)80009-3 Diego Nehab, André Maximo, Rodolfo S. Lima, and Hugues Hoppe

    work page doi:10.1016/s0146-664x(81)80009-3 1981

  7. [24]

    InProceedings of the 2011 SIGGRAPH Asia Conference(Hong Kong, China)(SA ’11)

    GPU-Efficient Recursive Filtering and Summed-Area Tables. InProceedings of the 2011 SIGGRAPH Asia Conference(Hong Kong, China)(SA ’11). Association for Computing Machinery, New York, NY, USA, Article 176, 12 pages. https://doi.org/10.1145/2024156.2024210 Shahan C. Nercessian

    arXiv 2011

  8. [26]

    https://doi.org/10.1109/TASSP.1984.1164348 Alan V

    Interpolated finite impulse response filters.IEEE Transactions on Acoustics, Speech, and Signal Processing32, 3 (1984), 563–570. https://doi.org/10.1109/TASSP.1984.1164348 Alan V. Oppenheim and Ronald W. Schafer. 2013.Discrete-Time Signal Processing(3rd ed.). Prentice Hall Press, USA. 532–550 pages. K.K. Parhi and D.G. Messerschmitt

    arXiv 1984

  9. [34]

    https: //api.semanticscholar.org/CorpusID:250627329 P

    Style Transfer of Audio Effects with Differentiable Signal Processing.ArXivabs/2207.08759 (2022). https: //api.semanticscholar.org/CorpusID:250627329 P. Stoica and T. Soderstrom

    Pith/arXiv arXiv 2022

  10. [35]

    The Steiglitz-McBride identification algorithm revisited–Convergence analysis and accuracy aspects.IEEE Trans. Automat. Control 26, 3 (1981), 712–717. https://doi.org/10.1109/TAC.1981.1102679 Andy Swanson

    arXiv 1981

  11. [36]

    https://www

    Latency and Its Effect on Performers. https://www. churchproduction.com/education/latency-and-its-affect-on-performers/. Accessed: 2025-04-30. A. Wang and J.O. Smith

    2025

  12. [40]

    Graph.39, 6, Article 188 (Nov

    RoboGrammar: graph grammar for terrain-optimized robot design.ACM Trans. Graph.39, 6, Article 188 (Nov. 2020), 16 pages. https://doi.org/10.1145/3414685.3417831 Haisen Zhao, Max Willsey, Amy Zhu, Chandrakana Nandi, Zachary Tatlock, Justin Solomon, and Adriana Schulz

    arXiv 2020

  13. [41]

    Co-optimization of design and fabrication plans for carpentry.ACM Transactions on Graphics (TOG)41, 3 (2022), 1–13

    2022

  14. [1973]

    https://doi.org/10

    Approximate design relationships for low-pass FIR digital filters.IEEE Transactions on Audio and Electroacoustics21, 5 (1973), 456–460. https://doi.org/10. 1109/TAU.1973.1162510 L. Rabiner, N. Graham, and H. Helms

    arXiv 1973

  15. [1974]

    https://doi.org/10.1109/TASSP.1974.1162558 Jonathan Ragan-Kelley, Andrew Adams, Sylvain Paris, Marc Levoy, Saman Amaras- inghe, and Frédo Durand

    Linear programming design of IIR digital filters with arbitrary magnitude function.IEEE Transactions on Acoustics, Speech, and Signal Processing22, 2 (1974), 117–123. https://doi.org/10.1109/TASSP.1974.1162558 Jonathan Ragan-Kelley, Andrew Adams, Sylvain Paris, Marc Levoy, Saman Amaras- inghe, and Frédo Durand

    arXiv 1974

  16. [1975]

    https://doi.org/10.1109/TASSP

    Optimum FIR digital filter implementations for decimation, interpolation, and narrow-band filtering.IEEE Transactions on Acoustics, Speech, and Signal Processing23, 5 (1975), 444–456. https://doi.org/10.1109/TASSP. 1975.1162719 Franklin C. Crow

    arXiv 1975

  17. [1981]

    https://doi.org/10.1016/0146-664X(81)90092-7 Y

    Fast filter transform for image processing.Computer Graphics and Image Processing16, 1 (1981), 20–51. https://doi.org/10.1016/0146-664X(81)90092-7 Y. Chan and R. Langford

    work page doi:10.1016/0146-664x(81)90092-7 1981

  18. [1982]

    https://doi.org/10.1109/TASSP.1982.1163946 Gaurav Chaurasia, Jonathan Ragan-Kelley, Sylvain Paris, George Drettakis, and Fredo Durand

    Spectral estimation via the high-order Yule-Walker equations.IEEE Transactions on Acoustics, Speech, and Signal Processing30, 5 (1982), 689–698. https://doi.org/10.1109/TASSP.1982.1163946 Gaurav Chaurasia, Jonathan Ragan-Kelley, Sylvain Paris, George Drettakis, and Fredo Durand

    arXiv 1982

  19. [1984]

    InProceedings of the 11th Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH ’84)

    Summed-area tables for texture mapping. InProceedings of the 11th Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH ’84). Association for Computing Machinery, New York, NY, USA, 207–212. https: //doi.org/10.1145/800031.808600 Rachid Deriche

    arXiv

  20. [1986]

    https://doi.org/10.1109/TCS.1986.1085930 Y.C

    Frequency-response masking approach for the synthesis of sharp linear phase digital filters.IEEE Transactions on Circuits and Systems33, 4 (1986), 357–364. https://doi.org/10.1109/TCS.1986.1085930 Y.C. Lim and Y. Lian

    arXiv 1986

  21. [1987]

    InPro- ceedings of the Second International Conference on Architectual Support for Pro- gramming Languages and Operating Systems(Palo Alto, California, USA)(AS- PLOS II)

    Superoptimizer: a look at the smallest program. InPro- ceedings of the Second International Conference on Architectual Support for Pro- gramming Languages and Operating Systems(Palo Alto, California, USA)(AS- PLOS II). Association for Computing Machinery, New York, NY, USA, 122–126. https://doi.org/10.1145/36206.36194 Naotake Masuda and Daisuke Saito

    arXiv

  22. [1988]

    In1988., IEEE International Symposium on Circuits and Systems

    Computationally efficient wideband filters with very narrow transition bands. In1988., IEEE International Symposium on Circuits and Systems. 2013–2017 vol.3. https://doi.org/10.1109/ISCAS.1988.15337 RawTherapee Development Team

    arXiv 2013

  23. [1989]

    Pipeline interleaving and parallelism in recursive digital filters. I. Pipelining using scattered look-ahead and decomposition. IEEE Transactions on Acoustics, Speech, and Signal Processing37, 7 (1989), 1099–1117. https://doi.org/10.1109/29.32286 PIXLS.US community

    work page doi:10.1109/29.32286 1989

  24. [1990]

    https://doi.org/10.1109/34

    Fast algorithms for low-level vision.IEEE Transactions on Pattern Analysis and Machine Intelligence12, 1 (1990), 78–87. https://doi.org/10.1109/34. 41386 R. Deriche and J. Abramatic

    doi:10.1109/34 1990

  25. [1993]

    https://doi.org/10.1109/82.219838 Karima Ma, Michael Gharbi, Andrew Adams, Shoaib Kamil, Tzu-Mao Li, Connelly Barnes, and Jonathan Ragan-Kelley

    The optimum design of one- and two-dimensional FIR filters using the frequency response masking technique.IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing40, 2 (1993), 88–95. https://doi.org/10.1109/82.219838 Karima Ma, Michael Gharbi, Andrew Adams, Shoaib Kamil, Tzu-Mao Li, Connelly Barnes, and Jonathan Ragan-Kelley

    work page doi:10.1109/82.219838 1993

  26. [1995]

    Signal Processing44, 2 (1995), 139–151

    Recursive implementation of the Gaussian filter. Signal Processing44, 2 (1995), 139–151. https://doi.org/10.1016/0165-1684(95)00020- E Ian T Young, Lucas J Van Vliet, and Michael Van Ginkel

    work page doi:10.1016/0165-1684(95)00020- 1995

  27. [1997]

    https: //doi.org/10.1109/78.599954 Ian T

    On fast FIR filters implemented as tail-canceling IIR filters.IEEE Transactions on Signal Processing45, 6 (1997), 1415–1427. https: //doi.org/10.1109/78.599954 Ian T. Young and Lucas J. van Vliet

    work page doi:10.1109/78.599954 1997

  28. [1998]

    In1998 IEEE Symposium on Advances in Digital Filtering and Signal Processing

    Weighted least squares IIR filter design with arbitrary magnitude and phase responses and specified stability margin. In1998 IEEE Symposium on Advances in Digital Filtering and Signal Processing. Symposium Proceedings (Cat. No.98EX185). 82–86. https://doi.org/10.1109/ADFSP.1998.685700 Yong Lim

    arXiv 1998

  29. [2002]

    IEEE Transactions on Signal Processing50, 11 (2002), 2798–2805

    Recursive gabor filtering. IEEE Transactions on Signal Processing50, 11 (2002), 2798–2805. Allan Zhao, Jie Xu, Mina Konaković-Luković, Josephine Hughes, Andrew Spielberg, Daniela Rus, and Wojciech Matusik

    2002

  30. [2011]

    In2011 18th IEEE International Conference on Image Processing

    Cosine integral images for fast spatial and range filtering. In2011 18th IEEE International Conference on Image Processing. IEEE, New York, NY, USA, 89–92. https://doi.org/10.1109/ICIP.2011.6116704 Zeev Farbman, Raanan Fattal, and Dani Lischinski

    arXiv 2011

  31. [2012]

    31, 4 (jul 2012), 32:1–32:12

    Decoupling Algorithms from Schedules for Easy Optimization of Image Processing Pipelines. 31, 4 (jul 2012), 32:1–32:12. G. Rajan, Y. Neuvo, and S.K. Mitra

    2012

  32. [2014]

    Graph.33, 4 (2014), 144–1

    Darkroom: compiling high-level image processing code into hardware pipelines.ACM Trans. Graph.33, 4 (2014), 144–1. https://doi.org/10.1145/2601097.2601174 Intel. 2020.Intel Integrated Performance Primitives. Technical Report. In- tel. https://www.intel.com/content/www/us/en/develop/documentation/ipp-dev- reference/top.html jordanthebamf

    arXiv 2014

  33. [2015]

    InProceedings of the 7th Conference on High-Performance Graphics(Los Angeles, California)

    Compiling High Performance Recursive Filters. InProceedings of the 7th Conference on High-Performance Graphics(Los Angeles, California). ACM, New York, NY, USA, 85–94. https://doi.org/10.1145/2790060.2790063 Joseph T. Colonel, Christian J. Steinmetz, Marcus Michelen, and Joshua D. Reiss

    arXiv

  34. [2017]

    A database of head- related transfer functions and morphological measurements.Journal of the Audio Engineering Society143 (October 2017). C. Steinmetz, Nicholas J. Bryan, and Joshua D. Reiss

    2017

  35. [2019]

    In7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9,

    The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks. In7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9,

    2019

  36. [2020]

    IneDAFx 2020 Proceedings of the International Conference on Digital Audio Effects

    Neural Parametric Equalizer Matching Using Differentiable Biquads. IneDAFx 2020 Proceedings of the International Conference on Digital Audio Effects. https://api.semanticscholar.org/CorpusID:226313578 Y. Neuvo, Dong Cheng-Yu, and S. Mitra

    2020

  37. [2021]

    InProceedings of the 22nd International Society for Music Information Retrieval Conference

    Synthesizer Sound Matching with Differ- entiable DSP. InProceedings of the 22nd International Society for Music Information Retrieval Conference. ISMIR, 428–434. https://doi.org/10.5281/zenodo.5624609 M.J. McDonnell

    work page doi:10.5281/zenodo.5624609

  38. [2022]

    https://www.apple

    GarageBand (Version 10.4.6) [Computer software]. https://www.apple. com/mac/garageband/. Franz Baader and Tobias Nipkow. 1998.Term rewriting and all that. Cambridge University Press, USA. Peter J Burt

    1998

  39. [2023]

    A Review of Differentiable Digital Signal Processing for Music & Speech Synthesis.ArXivabs/2308.15422 (2023). https://api.semanticscholar.org/CorpusID: 261276761 James Hegarty, John Brunhaver, Zachary DeVito, Jonathan Ragan-Kelley, Noy Cohen, Steven Bell, Artem Vasilyev, Mark Horowitz, and Pat Hanrahan

    Pith/arXiv arXiv 2023

  40. [2026]

    https://raw.pixls.us

    PIXLS.US Raw Image Database. https://raw.pixls.us. Public-domain (CC0) camera raw samples. Individual images referenced by their 62•Karima Ma, Andrew Adams, and Jonathan Ragan-Kelley getfile.phpURLs. Accessed 2026-07-21. L. Rabiner

    2026