SpectralNet: A Resolvent-Inspired Neural Architecture Based on Chernoff Approximations and Photonic Motivation

{"SpectralNet":[0],"is":[1,18,190,392,413,461],"a":[2,14,21,28,43,57,62,232,236,327,339,398,405],"digital":[3],"neural":[4],"architecture":[5],"motivated":[6],"by":[7,56,224,487],"the":[8,31,40,49,89,95,113,120,219,241,248,253,256,274,305,323,360,409,436,450,458,466,483,490,505,512],"Chernoff–Remizov":[9],"constructiveline":[10],"for":[11,123,194,211,433],"operator":[12,51],"evolution:":[13],"learnable":[15,414],"spectral":[16,446],"layer":[17],"read":[19],"as":[20,42,397],"discrete":[22],"small-stepfamily":[23],"F_phi(tau)":[24],"whose":[25],"composition":[26,329],"approximates":[27],"semigroup,":[29],"and":[30,53,61,94,207,244,269,355,380,457,471,509],"intermediate":[32],"statesare":[33],"aggregated":[34],"with":[35,227,277,387,442,449,469,511],"decaying":[36],"weights":[37],"modelled":[38],"on":[39,136,150,165,197,201,216,290,304,353],"resolvent":[41,251,410],"Laplace":[44],"transform.":[45],"Werestrict":[46],"ourselves":[47],"to":[48],"Fourier-diagonal":[50,275,324],"class":[52,300,325],"extend":[54],"it":[55],"low-rankcross-frequency":[58],"mixer":[59],"D+UV*":[60],"shift-based":[63],"spatial":[64,96],"block":[65],"S_{a,b,c}.":[66],"Main":[67],"result":[68,495],"(resource-matched":[69],"control":[70,468],"C.6,":[71],"three":[72],"datasets).":[73],"Under":[74],"one":[75],"training":[76],"recipe":[77],"anda":[78],"parameter":[79,367,411],"budget":[80],"of":[81,178,250,359,365,375,439,445],"~2.1–2.3":[82],"M":[83],"we":[84],"compare":[85],"two":[86,242],"evolution-step":[87],"families:":[88],"spectral-hybridstep":[90],"F^{-1}(D+UV*)F":[91],"∘":[92,100],"S_{a,b,c}":[93,101],"shift-rich":[97],"step":[98,172],"Pi_mix":[99],"(RMSB-R1).Top-1":[102],"accuracy":[103,175,183],"(mean":[104],"±":[105,128,133,143,147,157,162,281,286],"sample":[106],"std":[107],"over":[108,115,185],"seeds;":[109],"each":[110],"run":[111],"reports":[112],"maximum":[114],"epochs":[116],"ofthe":[117],"test-set":[118],"accuracy,":[119],"same":[121,513],"rule":[122],"every":[124,500],"model):":[125],"RMSB-R1":[126,221],"89.87":[127],"0.20":[129],"%":[130,135,145,149,159,164,283,288,312,316,321],"vsspectral-hybrid":[131],"80.32":[132],"0.38":[134],"CIFAR-10":[137],"(5":[138,152],"seeds,":[139,153,167],"+9.55":[140],"pp),":[141,155,206],"62.90":[142],"0.24":[144],"vs52.32":[146],"0.43":[148],"CIFAR-100":[151],"+10.58":[154],"95.90":[156],"0.09":[158],"vs":[160,319],"89.22":[161],"0.17":[163],"SVHN(3":[166],"+6.68":[168],"pp).":[169],"The":[170],"spectral-hybrid":[171],"loses":[173],"less":[174],"under":[176],"corruption":[177],"thetest":[179],"images:":[180],"its":[181,366],"mean":[182,268],"drop":[184,308],"severity":[186],"levels":[187],"(\\"AUC":[188],"drop\\")":[189],"smaller":[191],"by21.2/16.6":[192],"pp":[193,210,226,432],"Gaussian":[195,214],"blur":[196,296,475],"CIFAR-10/100":[198],"but":[199],"equal":[200],"SVHN":[202,474],"(Δ":[203],"=":[204,477],"−0.14":[205,478],"smallerby":[208],"15.2/14.3/5.6":[209],"additive":[212],"white":[213],"noise":[215],"CIFAR-10/100/SVHN.":[217],"Within":[218],"sameprotocol":[220],"exceeds":[222],"ResNet-18":[223],"+3.81/+7.61/+2.34":[225],"~5×":[228],"fewer":[229],"parameters;":[230],"thisis":[231],"within-protocol":[233],"observation,":[234],"not":[235,246,302,337,357,404],"general":[237],"claim.":[238],"C.6":[239,467],"compares":[240],"evolution-stepfamilies":[243],"does":[245,301,336],"isolate":[247],"contribution":[249],"aggregation;":[252],"only":[254],"ablationof":[255],"aggregation":[257,437],"(single-seed":[258],"SpectralNet-S,":[259],"corrected":[260],"in":[261,298,415,489,507],"v1.1)":[262],"shows":[263,454],"no":[264,455,494],"measurabledifference":[265],"between":[266],"resolvent,":[267],"last":[270],"aggregation.":[271],"Further":[272],"results:":[273],"SpectralNet-B":[276],"GELU":[278],"reaches":[279],"76.50":[280],"0.16":[282],"(n=2)and":[284],"76.67":[285],"0.18":[287],"(n=3)":[289],"CIFAR-10,":[291],"above":[292],"ShuffleNetV2":[293],"(74.73":[294],"%);":[295],"robustness":[297],"thebase":[299],"depend":[303],"activation":[306],"(relative":[307],"at":[309],"σ=2:":[310],"43.9":[311],"GELU-B":[313],"n=2":[314],"≈44.5":[315],"PhysicalAct-B":[317],"n=1":[318],"≈63":[320],"ResNet-18);":[322],"has":[326],"limiteduseful":[328],"horizon":[330],"(2–3":[331],"evolution":[332],"steps).":[333],"This":[334],"paper":[335],"claim":[338,460],"hardware":[340],"photonic":[341],"realization.":[342],"Remizov's":[343,384],"theorems":[344,385],"are":[345,423,497],"used":[346],"asmotivation":[347],"only:":[348],"they":[349],"concern":[350],"fixed-coefficient":[351],"generators":[352],"UC_b(R)":[354],"do":[356],"assertconvergence":[358],"trained":[361],"finite-depth":[362],"network":[363],"or":[364],"derivatives.":[368],"Version":[369],"1.1":[370],"(September":[371],"2026):":[372],"text":[373],"revision":[374],"v1.0":[376,440,459],"(DOI":[377],"10.5281/zenodo.19452600);":[378],"sameexperiments":[379],"archive.":[381],"Changes:":[382],"(i)":[383],"restated":[386],"hypotheses":[388],"—":[389,448,499],"onlyTheorem":[390],"6":[391],"shift-based;":[393],"(ii)":[394],"RemizovShiftLayer":[395],"described":[396],"2-D":[399],"four-shift":[400],"analogue":[401],"ofTheorem":[402],"6,":[403],"direct":[406],"implementation;":[407],"(iii)":[408],"λ":[412],"thereleased":[416],"code":[417],"(v1.0:":[418],"fixed);":[419],"(iv)":[420],"reported":[421],"accuracies":[422],"best-epoch":[424],"test":[425],"accuracies,":[426],"novalidation":[427],"split":[428],"(effect":[429],"≤":[430],"0.32":[431],"C.6);":[434],"(v)":[435],"ablation":[438],"comparedruns":[441],"different":[443],"numbers":[444],"modes":[447],"matched":[451],"baseline":[452],"resolventaggregation":[453],"advantage,":[456],"withdrawn;":[462],"(vi)":[463],"abstract":[464],"restatedaround":[465],"metric":[470],"protocol;":[472],"(vii)":[473],"Δ":[476],"pp.":[479],"All":[480],"numbersre-derived":[481],"from":[482],"primary":[484],"experiment":[485],"outputs":[486],"scripts/verify_w1_claims.py":[488],"coderepository":[491],"(71":[492],"checks);":[493],"files":[496],"distributed":[498],"table":[501],"can":[502],"be":[503],"regeneratedwith":[504],"launchers":[506],"scripts/":[508],"checked":[510],"script.":[514],"Files:":[515],"main_en.pdf,main_ru.pdf":[516],"(Russian":[517],"version).":[518],"Code:":[519],"https://github.com/shpital/spectralnet":[520],"(tag":[521],"paper-v1.1).":[522]}

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-09-15
DOI
https://doi.org/10.5281/zenodo.22770171
Primary Topic
Neural Networks and Reservoir Computing
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
preprint

SpectralNet: A Resolvent-Inspired Neural Architecture Based on Chernoff Approximations and Photonic Motivation

Sergey Shpital
Zenodo (CERN European Organization for Nuclear Research)
Neural Networks and Reservoir Computing
preprint

SpectralNet: A Resolvent-Inspired Neural Architecture Based on Chernoff Approximations and Photonic Motivation

Sergey Shpital
preprint en

Abstract

SpectralNet is a digital neural architecture motivated by the Chernoff–Remizov constructiveline for operator evolution: a learnable spectral layer is read as a discrete small-stepfamily F_phi(tau) whose composition approximates a semigroup, and the intermediate statesare aggregated with decaying weights modelled on the resolvent as a Laplace transform. Werestrict ourselves to the Fourier-diagonal operator class and extend it by a low-rankcross-frequency mixer D+UV* and a shift-based spatial block S_{a,b,c}. Main result (resource-matched control C.6, three datasets). Under one training recipe anda parameter budget of ~2.1–2.3 M we compare two evolution-step families: the spectral-hybridstep F^{-1}(D+UV*)F ∘ S_{a,b,c} and the spatial shift-rich step Pi_mix ∘ S_{a,b,c} (RMSB-R1).Top-1 accuracy (mean ± sample std over seeds; each run reports the maximum over epochs ofthe test-set accuracy, the same rule for every model): RMSB-R1 89.87 ± 0.20 % vsspectral-hybrid 80.32 ± 0.38 % on CIFAR-10 (5 seeds, +9.55 pp), 62.90 ± 0.24 % vs52.32 ± 0.43 % on CIFAR-100 (5 seeds, +10.58 pp), 95.90 ± 0.09 % vs 89.22 ± 0.17 % on SVHN(3 seeds, +6.68 pp). The spectral-hybrid step loses less accuracy under corruption of thetest images: its mean accuracy drop over severity levels ("AUC drop") is smaller by21.2/16.6 pp for Gaussian blur on CIFAR-10/100 but equal on SVHN (Δ = −0.14 pp), and smallerby 15.2/14.3/5.6 pp for additive white Gaussian noise on CIFAR-10/100/SVHN. Within the sameprotocol RMSB-R1 exceeds ResNet-18 by +3.81/+7.61/+2.34 pp with ~5× fewer parameters; thisis a within-protocol observation, not a general claim. C.6 compares the two evolution-stepfamilies and does not isolate the contribution of resolvent aggregation; the only ablationof the aggregation (single-seed SpectralNet-S, corrected in v1.1) shows no measurabledifference between resolvent, mean and last aggregation. Further results: the Fourier-diagonal SpectralNet-B with GELU reaches 76.50 ± 0.16 % (n=2)and 76.67 ± 0.18 % (n=3) on CIFAR-10, above ShuffleNetV2 (74.73 %); blur robustness in thebase class does not depend on the activation (relative drop at σ=2: 43.9 % GELU-B n=2 ≈44.5 % PhysicalAct-B n=1 vs ≈63 % ResNet-18); the Fourier-diagonal class has a limiteduseful composition horizon (2–3 evolution steps). This paper does not claim a hardware photonic realization. Remizov's theorems are used asmotivation only: they concern fixed-coefficient generators on UC_b(R) and do not assertconvergence of the trained finite-depth network or of its parameter derivatives. Version 1.1 (September 2026): text revision of v1.0 (DOI 10.5281/zenodo.19452600); sameexperiments and archive. Changes: (i) Remizov's theorems restated with hypotheses — onlyTheorem 6 is shift-based; (ii) RemizovShiftLayer described as a 2-D four-shift analogue ofTheorem 6, not a direct implementation; (iii) the resolvent parameter λ is learnable in thereleased code (v1.0: fixed); (iv) reported accuracies are best-epoch test accuracies, novalidation split (effect ≤ 0.32 pp for C.6); (v) the aggregation ablation of v1.0 comparedruns with different numbers of spectral modes — with the matched baseline resolventaggregation shows no advantage, and the v1.0 claim is withdrawn; (vi) abstract restatedaround the C.6 control with metric and protocol; (vii) SVHN blur Δ = −0.14 pp. All numbersre-derived from the primary experiment outputs by scripts/verify_w1_claims.py in the coderepository (71 checks); no result files are distributed — every table can be regeneratedwith the launchers in scripts/ and checked with the same script. Files: main_en.pdf,main_ru.pdf (Russian version). Code: https://github.com/shpital/spectralnet (tag paper-v1.1).

Zenodo (CERN European Organization for Nuclear Research)
Peace, Justice and strong institutions
Neural Networks and Reservoir Computing
DOIOpen Access PDFOpenAlex
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.