7
Group Convolution on Homogeneous Spaces
Convolutionalneuralnetworksrelyonmatchinginputfeatureswith
appropriately-transformedfilterparameters;thisideaextendswell
beyond shift-equivariance on grids.
Wecandefinearichfamilyofgroupconvolutionswhichapplyamore
generalfiltertransformation,followedbyaninnerproductwiththe
features over the domain—so long as the input domain is homogeneous.
Forshift-equivariantfunctionsongrids,thedomainandgrouphave
identicalstructure—constructingthemoregeneralcaserequirescarefully
keeping track of the structure within the group.
This allows ustoconstruct convolutionalneuralnetworks over spheres,
DNA sequences, 3D medical scans, and many more domains.
Ourdiscussionofgridshighlightedhowshiftsandconvolutionsareinti-
matelyconnected:convolutionsarelinearshift-equivariant
1
operations,and
vice versa, any shift-equivariantlinear operator is a convolution. Furthermore,
shift operatorscanbe jointlydiagonalisedby theFouriertransform.As itturns
out,thisispartofa farlargerstory:bothconvolutionandtheFouriertransform
can be defined for any group of symmetries that we can sum or integrate over.
ConsidertheEuclideandomainΩ=R.Wecanunderstandtheconvolution
asapatternmatchingoperation:wematchshiftedcopiesofafilterθ(u)with
aninputsignalx(u).Thevalueoftheconvolution(x⋆θ)(u)atapointuisthe
inner product of the signal x with the filter shifted by u,
(x⋆θ)(u)=x,S
u
θ=
Z
R
x(v)θ(u+v)dv.
2
Notethatinthis caseu isboth apointonthedomainΩ=Randalsoan element
ofthetranslation group,whichwecanidentifywiththedomainitself,G=R.
Wewillnowshowhowtogeneralise thisconstruction,bysimply replacingthe
translation group by another group Gacting on Ω.
194Chapter 7
7.1Domain
As discussed in Chapter3, the action ofthe group Gon thedomain Ω induces
a representationρof Gon thespace ofsignals X(Ω)via ρ(g)x(u)=x(g
–1
u). In
theaboveexample,Gisthetranslationgroupwhoseelementsactbyshifting
thecoordinates,u+v,whereasρ(g)istheshiftoperatoractingonsignalsas
(S
v
x)(u)=x(uv). Finally,inorder toapply afilterto thesignal, weinvokeour
assumption of X(Ω) being a Hilbert space, with an inner product
x,θ=
Z
Ω
x(u)θ(u)du,
3
whereweassumed, forthesakeofsimplicity,scalar-valuedsignals,X(Ω,R).
Having thus defined how to transformsignals andmatch themwith filters,we
can define the group convolution for signals on Ω,
(x⋆θ)(g)=x,ρ(g)θ=
Z
Ω
x(u)θ(g
–1
u)du.(7.96)
Note thatx⋆θtakesvalueson theelements gof ourgroup Grather than points
on thedomainΩ.Hence,the nextlayer, whichtakesx⋆θasinput, shouldact
on signals defined on to the group G, a point we will return to shortly.
JustlikehowthetraditionalEuclideanconvolutionisshift-equivariant,the
moregeneralgroupconvolutionisG-equivariant. Thekeyobservation isthat
matching thesignalx withag-transformed filterρ(g)θis thesameas matching
theinversetransformedsignalρ(g
–1
)xwiththeuntransformedfilterθ.Math-
ematically, thiscan beexpressed as x,ρ(g)θ=ρ(g
–1
)x,θ.With thisinsight,
G-equivariance of the group convolution(Equation 7.96) follows immediately
fromitsdefinitionandthedefiningpropertyρ(h
–1
)ρ(g)=ρ(h
–1
g)ofgroup
representations,
(ρ(h)x⋆θ)(g)=ρ(h)x,ρ(g)θ=x,ρ(h
–1
g)θ=ρ(h)(x⋆θ)(g).
TheG-equivariantgroupconvolutioncanbeseenasa "generator"of models:it
can provide a recipe for constructingG-equivariant neural networks given any
suitable G. Inordertoeventuallyunderstandhow toground thisintoconcrete
model equations, we need to look at specific examples.
7.1.1Grid convolution
Thecaseofshiftequivarianceoverthe(one-dimensional)gridwehavestud-
iedthroughoutChapter6isobtainedwiththechoiceΩ=Z
n
={0,...,n–1}
andthecyclicshiftgroupG=Z
n
.Thegroupelementsinthiscasearecyclic
shiftsofindices,i.e.,anelementg∈Gcanbeidentifiedwithsomeu
{0,...,n–1}suchthatg.v=(vu)modn,whereastheinverseelementis
g
–1
.v=(v+u)modn.Importantly,inthisexampletheelementsofthegroup
Groups195
Figure 7.1
Left: Cosmic microwave background radiation, captured by thePlanck space observa-
tory, is asignal on S
2
. Right: Theaction of the specialorthogonal group, SO(3), onthe
sphere, S
2
. Threetypes of rotation arepossible; SO(3) isa three-dimensional manifold.
(shifts)arealsoelementsofthedomain(indices).Wethuscan,withsome
abuseofnotation,identifythetwostructures(i.e.,Ω=G);ourexpressionfor
the group convolution in this case
(x⋆θ)(g)=
n–1
X
v=0
x
v
θ
g
–1
.v
,
leads to the familiar convolution
4
(x⋆θ)
u
=
n–1
X
v=0
x
v
θ
v+umod n
.
7.1.2Spherical convolution
Nowconsiderthetwo-dimensionalsphereΩ=S
2
withthegroupofrota-
tions,thespecialorthogonalgroupG=SO(3).Whilechosenforpedagogical
reasons,thisexampleisactuallyverypracticalandarisesinnumerousappli-
cations.Inastrophysics,forexample,observationaldataoftennaturallyhas
sphericalgeometry(Figure7.1).Thesamecanbesaidofanytaskinvolv-
ingweatherprediction(Lametal.2023).Furthermore,sphericalsymmetries
are very important in applications in chemistry whenmodeling molecules and
trying topredict theirproperties, e.g. forthe purposeof virtualdrug screening.
Representingapointonthesphereasathree-dimensionalunitvectoru:
u=1,theactionofthegroupcanberepresentedasa3×ばつ3orthogonalmatrix
Rwithdet(R)=1.Thesphericalconvolution canthusbewrittenastheinner
196Chapter 7
product between the signal and the rotated filter,
(x⋆θ)(R)=
Z
S
2
x(u)θ(R
–1
u)du.
Thefirstthingtonoteisthannowthegroupisnotidenticaltothedomain:
thegroupSO(3)isaLiegroupthatisinfactathree-dimensionalmanifold,
whereasS
2
isatwo-dimensionalone.Consequently,inthiscase,unlikethe
previous example, the convolution is a function on SO(3) rather than on Ω.
Thishasimportantpracticalconsequences:inourGeometric DeepLearning
blueprint,we concatenatemultipleequivariantmaps("layers"indeeplearning
jargon)by applyinga subsequentoperator to theoutput ofthe previous one. In
the caseof translations, wecan applymultiple convolutions insequence, since
theiroutputsarealldefinedonthesamedomainΩ.Inthegeneralsetting,
sincex⋆θisafunctiononGratherthanonΩ,wecannotuseexactlythe
sameoperationsubsequently—it meansthatthenextoperationhasto dealwith
signalsonG,i.e.x∈X(G).Ourdefinitionofgroupconvolutionallowsthis
case:wetakeasdomainΩ=Gactedon byGitselfviathegroupaction(g,h)7→
ghdefinedbythecompositionoperationofG.Thisyieldstherepresentation
ρ(g)actingonx∈X(G)by(ρ(g)x)(h)=x(g
–1
h)
5
.Justlikebefore,theinner
productisdefinedbyintegratingthepoint-wiseproductofthesignalandthe
filteroverthedomain,whichnowequalsΩ=G.Inourexampleofspherical
convolution, a second layer of convolution would thus have the form
((x⋆θ)⋆φ)(R)=
Z
SO(3)
(x⋆θ)(Q)φ(R
–1
Q)dQ.
7.1.3Limitations
Since convolutions involve inner products, that inturn require integrating over
the domainΩ,wecan onlyuseiton domainsΩ thatare small(inthediscrete
case) or low-dimensional (in the continuous case).
Forinstance,wecanuseconvolutionsontheplaneR
2
(twodimensional)
orspecialorthogonalgroupSE(3)(threedimensional),oronthefinitesetof
nodesofagraph(n-dimensional).Itmightthenbetemptingtoconstructa
highlyexpressivegraphneuralnetworkby performingthiskindofconvolution
directly onthe groupofpermutations S
n
. We can,for example,first transform
features from a set of nodes Vinto S
n
:
(x⋆θ)(σ)=
X
u∈V
x(u)θ(σ
–1
(u)),(7.97)
and then continue transforming on S
n
as follows:
((x⋆θ)⋆φ)(σ)=
X
σ
S
n
(x⋆θ)(σ
)φ(σ
–1
σ
),(7.98)
Groups197
θ
1
θ
3
θ
2
φ
312
φ
132
φ
321
φ
231
φ
123
φ
213
Inputnodes,V
u
1
u
2
u
3
S
3
(Lifted)
σ
(123)
σ
(213)
σ
(132)
σ
(312)
σ
(231)
σ
(321)
S
3
(Convolved)
σ
(123)
σ
(213)
σ
(132)
σ
(312)
σ
(231)
σ
(321)
Figure 7.2
Leveragingthegroupconvolutional framework toconstructalearnableS
3
-equivariant
transformationoverthree-nodegraphs,perEquations7.97–7.98.Thefirstlayermaps
nodefeatures inX(V) directlyto permutationfeatures inX(S
3
),via parametersθ:V→
R;thesecondlayermapsX(S
3
)→X(S
3
)viaparametersφ:S
3
R.Whileelegant
and spanning a rich class of permutation-equivariant graph models, it quickly becomes
unwieldy at larger |V|, and does not easily transfer across graphs of different sizes.
whereσ,σ
S
n
arepermutationsandispermutationcomposition(see
Figure 7.2 for an example over three nodes and six permutations).
However,inpractice,suchalayercannotbeconstructedonanybutthe
smallestofgraphs,becausethepermutationgroupS
n
hasn!elements,and
thelayeraboverequiresstoringafeaturevector,(x⋆θ)(σ),foreachofthose
elements.Whilesuchconstructionsareindeedimpractical,itisinterestingto
ponderthatthereexistsanextremelyrichfamilyofpermutation-equivariant
modelsactingdirectlyonthepermutationgroupinthisway,encompassing
layers of possibly significant amounts of expressive power.
Similarly,integratingoverhigher-dimensionalgroupsliketheaffinegroup
(containing translations,rotations, shearingand scaling,for atotal of6 dimen-
sions)isnotfeasibleinpractice.Nevertheless,aswehaveseeninChapter5,
we can still build equivariantconvolutionsfor large groupsGby workingwith
signalsdefinedonlow-dimensionalspacesΩonwhichGacts.Indeed,itis
possibletoshowthatanyequivariantlinearmapf:X(Ω)→X(Ω
)between
198Chapter 7
twodomainsΩ,Ω
canbewrittenasageneralisedconvolutionsimilartothe
group convolution discussed here.
Second,wenotethatthe Fourier transformwederivedinChapter6fromthe
shift-equivariancepropertyoftheconvolution canalsobeextended toamore
generalcasebyprojectingthesignalontothematrixelementsofirreducible
representations of G.
Finally,wepointtotheassumptionthathassofarunderpinnedourdiscus-
sioninthisChapter:whetherΩwasagrid,plane,orthesphere,wecould
transformeverypointintoanyotherpoint,intuitivelymeaningthatallthe
pointsonthedomain"lookthesame."AdomainΩwithsuchpropertyis
calledahomogeneousspace,whereforanyu,vΩthereexistsg∈Gsuch
that g.u=v
6
. In future Chapters, we will try to relax this assumption.
7.2Model
Asdiscussedthusfar,wecangeneralisetheconvolutionoperationfromsignals
onaEuclideanspacetosignalsonanyhomogeneousspaceΩ acteduponby
agroupG.ByanalogytotheEuclideanconvolution, whereatranslated filter
ismatchedwiththesignal,theideaofgroupconvolutionistomove thefilter
around the domainusing thegroup action,e.g. by rotatingand translating.By
virtueofthetransitivityofthegroupaction,wecanmovethefiltertoany
positiononΩ.Inthissection,wewilldiscussseveralconcreteexamplesof
thegeneralideaofgroupconvolution,includingimplementationaspectsand
architectural choices.
7.2.1Discrete group convolution
WebeginbyconsideringthecasewherethedomainΩaswellasthegroup
Garediscrete.Asourfirstexample,weconsidermedicalvolumetricimages
representedassignalsofon3Dgridswithdiscretetranslationandrotation
symmetries.Thedomainisthe3DcubicalgridΩ=Z
3
andtheimages(e.g.
MRIorCT3Dscans)aremodelledasfunctionsx:Z
3
R,i.e.x∈X(Ω).
Although,inpractice,suchimageshavesupportonafinitecuboid[W]×ばつ[H]×ばつ
[D]Z
3
,weinsteadprefertoviewthemasfunctionsonZ
3
withappropri-
atezeropadding.Asoursymmetry,weconsiderthegroupG=Z
3
O
h
of
distance-andorientation-preservingtransformationsonZ
3
.Thisgroupcon-
sistsoftranslations(Z
3
)andthediscreterotationsO
h
generatedby90degree
rotations about the three axes (see Figure 7.3).
As oursecondexample, we considerDNA
7
sequences madeupof fourlet-
ters:C,G,A,andT.Thesequencescanberepresentedonthe1DgridΩ=Z
assignalsx:ZR
4
,whereeachletterisone-hotcodedinR
4
.Naturally,we
have a discrete 1D translation symmetry on the grid, but DNA sequences have
Groups199
Figure 7.3
A 3×ばつ3 filter, rotated byall 24 elements ofthe discrete rotationgroup O
h
, generatedby
90-degreerotationsabout the verticalaxis (red arrows),and 120-degree rotationsabout
a diagonal axis (blue arrows).
anadditionalinterestingsymmetry.ThissymmetryarisesfromthewayDNA
is physically embodiedas adoublehelix, andthe way itis readby themolec-
ularmachineryofthecell.Eachstrandofthedoublehelixbeginswithwhat
iscalledthe5
-endandendswitha3
-end,withthe5
ononestrandcom-
plementedbya3
ontheotherstrand.Inotherwords,thetwostrandshave
anoppositeorientation.SincetheDNAmoleculeisalwaysreadoffstarting
atthe5
-end,butwedonotknowwhichone,asequencesuchasACC-
CTGGisequivalenttothereversedsequencewitheachletterreplacedbyits
complement,CCAGGGT. Thisiscalledreverse-complement symmetryofthe
lettersequence,andisdepictedinFigure7.4.Wethushavethetwo-element
groupZ
2
={0,1}correspondingtotheidentity0andreverse-complement
transformation 1(andcomposition 1+1=0mod2). The fullgroupcombines
translations and reverse-complement transformations.
Inthisdiscretecase,thepreviouslydefinedgroupconvolution(Equation
7.96) is given as the following inner product:
(x⋆θ)(g)=
X
uΩ
x
u
ρ(g)θ
u
,(7.99)
between the (single-channel) input signal x and a filter θtransformed by g∈G
via ρ(g)θ
u
=θ
g
–1
u
, andtheoutput x⋆θisafunction onG.Note that,sinceΩ is
discrete, we have replaced the integral from Equation 7.96 by a sum.
200Chapter 7
3’
5’5’
3’
C
G
A
T
T
A
T
A
C
G
C
G
T
A
T
A
G
C
G
C
G
C
T
A
Figure 7.4
A schematic of the DNA’s double helix structure, with the two strands coloured in blue
and red.Notehow the sequencesinthe helicesarecomplementaryand readinreverse
(from 5’ to 3’).
7.2.2Transform + Convolve approach
Wewillshowthatthediscretegroupconvolution canbeimplementedintwo
steps:afiltertransformationstep,andatranslationalconvolutionstep.The
filtertransformationstepconsistsofcreatingrotated(orreverse-complement
transformed)copiesofa basicfilter, whilethetranslationalconvolution isthe
sameasinstandardCNNsandthusefficientlycomputableonhardwaresuch
asGPUs.To seethis,notethat,inbothof ourexamples, wecan writeageneral
transformationg∈Gasatransformationh∈H(e.g.arotationorreverse-
complement transformation)followed bya translationkZ
d
, i.e.g=kh (with
juxtapositiondenotingthecompositionofthegroupelementskandh).By
properties of the group representation, we have ρ(g)=ρ(kh)=ρ(k)ρ(h). Thus,
(x⋆θ)(kh)=
X
uΩ
x
u
ρ(k)ρ(h)θ
u
=
X
uΩ
x
u
(ρ(h)θ)
uk
(7.100)
Werecognisethelastequationasthestandard(planarEuclidean)convolu-
tionofthesignalxandthetransformedfilterρ(h)θ.Thus,toimplement
groupconvolutionforthesegroups,wetakethecanonicalfilterθ,create
transformedcopiesθ
h
=ρ(h)θforeachh∈H(e.g.eachrotationhO
h
or
reverse-complement DNA symmetryhZ
2
),andthenconvolvex witheachof
these filters: (x⋆θ)(kh)=(x⋆θ
h
)(k). For both of our examples, the symmetries
acton filtersbysimplypermuting thefiltercoefficients, asshownin Figure7.3
fordiscreterotations.Hence,theseoperationscanbeimplementedefficiently
using an indexing operation with pre-computed indices.
Whilewedefinedthefeaturemapsthatareproducedbythegroupconvo-
lutionx⋆θasfunctionsonG,thefactthatwecansplitanyg∈Gintog=hk
meansthatwecanalsothinkofthemasastackofEuclideanfeaturemaps
Groups201
(sometimes called orientation channels),with one feature map per filter trans-
formation /orientationk. Forinstance,inour firstexamplewewouldassociate
toeachfilterrotation(eachnodeinFigure7.3)afeaturemap,whichisobtained
byconvolving(inthetraditionaltranslationalsense)therotatedfilter.These
feature mapscanthusstillbe storedasaW ×ばつH ×ばつCarray,wherethe number
ofchannelsCequalsthenumberofindependentfilterstimesthenumberof
transformations h∈H(e.g. rotations).
Aspreviouslyshown,thegroupconvolutionisequivariant:(ρ(g)x)⋆θ=
ρ(g)(x⋆θ). Whatthis means in termsof orientation channels is that,under the
action ofh, eachorientation channelistransformed, andthe orientationchan-
nelsthemselvesarepermuted.Forinstance,ifweassociateoneorientation
channelpertransformationinFigure7.3andapplyarotationby90degrees
aboutthez-axis(correspondingtotheredarrows),thefeaturemapswillbe
permuted as shown by the red arrows.
Thisdescriptionmakesitclearthatagroupconvolutionalneuralnetwork
bearsmuchsimilaritytoatraditionalCNN.Hence,manyofthenetworkdesign
patternsdiscussedinChapter6,suchasresidualnetworks,canbeusedwith
group convolutions as well.
7.2.3Spherical CNNs via the Fourier domain
Forthe continuoussymmetrygroup ofthesphere thatwe saw inSection7.1.2,
itispossibletoefficientlyimplementtheconvolutioninthespectraldomain,
usingtheappropriateFouriertransform(weremindthereaderthatthecon-
volutiononS
2
isafunctiononSO(3),henceweneedtodefinetheFourier
transformonboththesedomainsinordertoimplementmulti-layerspherical
CNNs).Sphericalharmonicsareanorthogonalbasisonthe2Dsphere,anal-
ogoustotheclassicalFourierbasisofcomplexexponential.Onthespecial
orthogonal group,the Fourier basisis known as theWignerD-functions. Both
of these bases find wide applications in quantum mechanics and chemistry.
Inbothcases, theFouriertransforms(coefficients)arecomputedastheinner
product withthebasisfunctions,andan analogyoftheConvolution Theorem
holds: one can compute the convolution in the Fourier domain as the element-
wiseproductoftheFouriertransforms. Furthermore,FFT-likealgorithms exist
for theefficient computation ofthe Fourier transformonS
2
and SO(3).Using
thisapproach,Cohenetal.(2018)wereabletobuildthefirstefficientspherical
CNN; we defer to their paper for specific details.
7.3Case Study: TacticAI
Asmightbeexpected,groupconvolutionalnetworksmanifestinawidevariety
of interesting domains, but perhaps the reader might still find it surprising that
202Chapter 7
5
6
4
1
3
2
5
6
4
1
3
2
5
6
4
1
3
2
5
6
4
1
3
2
5
6
4
1
3
2
e
↔↕
x
4
x
5
x
6
x
1
x
2
x
3
x
4
x
5
x
6
x
1
x
2
x
3
x
6
x
5
x
4
x
3
x
2
x
1
x
6
x
5
x
4
x
3
x
2
x
1
X
e
X
X
X
↔↕
G
X
e
X
G
X
X
e
G
X
X
↔↕
G
X
↔↕
X
H
Figure 7.5
Illustrationof asingle layerof TacticAI’s group-equivariantneural network.Fora given
cornerkicksituation(herevisualisedusingonlysixplayersforclarity),TacticAIfirst
generates four views, corresponding to all four possible corner kick locations (e.g., X
fortheverticallyreflectedcorner).Then,agraphneuralnetwork,G,processescarefully
chosenpairsofviews,inawaythatfollowstheequivarianceconstraint.Lastly,allof
the computedview-pair representationsareaggregated toproducelatentview features
(e.g. H
for the vertically-reflected view).
theyhaveseennotableuseinassociationfootball
8
analytics.Toroundoff
thegroupsChapter,weprovideatargetedoverviewoftheresultingTacticAI
method,developedincollaborationwithLiverpoolFC(Wangetal.2024).
Rather than surveying the entire architecture and the tasks it was deployed on,
weparticularlyfocusonitsgroup-equivariantaspects(seealsoFigure7.5),
along with meaningful context on why these aspects were called for.
ProblemsetupTacticAIisasystemcapableofpredictiveandgenerative
modellingovervarioustacticalsetupsinfootball.Thetacticalsetupinputis
representedasagraphof22nodes—oneforeachplayer—witheachnode’s
features,x
u
R
k
,comprisingbothspatial(currentpositionandvelocity)and
physical(player height,weightandball possession)informationaboutthe cor-
responding player. It is assumed thatall pairs of players are connected to each
other,allowingthemodeltoinferthemostimportantconnectionsautomati-
cally(aswediscussedinChapter5).Thesystemneedstotheneitherpredict
Groups203
12
34
e
43
21
21
43
34
12
↔↕
Figure 7.6
TheCayleygraphofthedihedralgroupD
2
={e,,,↔↕},organising allofitsindi-
vidual elementsandtheiractionona domainspanningthefourcornersofa rectangle.
Notethatthearrowsarebidirectional,sinceeachD
2
groupactionisitsowninverse.
Further,notethatcomposingbothhorizontalandverticalreflectionsisequivalenttoa
180
rotation.
theoutcomeofafutureevent(e.g.whowillmakenextcontactwiththeball—a
node classification task,or willa shot betaken—a graphclassification task) or
generate novel setups (that, e.g., modulate the likelihood a shot will be taken).
As itisoftentricky tomeaningfullyinfluenceopen-playtacticsinfootball,
TacticAIfocussesallofitsattentiononmodellingoutcomesinsetpieces,
duringwhichthegameiseffectivelyfrozen.Specifically,cornerkicksetups
weretargeted,astheyoccurreasonablyfrequently,startfromarigidposi-
tion,andofferimmediategoal-scoringpotential.Further,cornerkicktactics
areoftendeterminedwellinadvanceofindividualmatches,allowingfora
cleaner window for TacticAI to influence coaching decisions.
AnunexpectedsymmetryThedecisiontofocusoncornerkicks alsobrings
with itan undesirabletradeoff—theyare simplynot extremelyfrequent, lead-
ingtorelativelymodestdatasetsizes.Indeed,eventhoughTacticAIwastrained
overseveralseasons’worthof PremierLeaguedata,only9,693uniquetrain-
ingexampleswereextractablefromthisdata—afarcryfromthelargescale
datasetsusedformanyofthecasestudiespreviouslydiscussedinthisbook.
Further coupled with thescarcity of certain target events (such as shots, which
are relatively rare), the quantity of useful signal is significantly reduced.
204Chapter 7
Atthislevelofdataavailability,exploitingsymmetriesarisesasanatural
approach for optimising theextent to which the provided data isutilised. That
said,itwasnot immediatelyobviouswheresuchsymmetriescouldcomefrom.
Owingto adirect suggestionfromthe LiverpoolFC collaborators
9
,it wascon-
cludedthat,ifonevariesthespecificcornerinwhichthesetpieceisbeing
taken(outofthefourpossiblecornersofthepitch),theoutcomeswouldremain
approximately equivariant.
The symmetrygroupwhich governschanging cornerpositionsin awaythat
preservestherelativepositioning ofallotherpointsisthedihedralgroup
10
,G=
D
2
.Thisgroupcomprisesonlyfourelements:D
2
={e,,,↔↕},enumerat-
ingfour possibletransformations: identity, verticalflip, horizontalflip,vertical
andhorizontalflip(180
rotation).ItisfullydescribedbyFigure7.6;note
thateachoperationisitsown inverse, i.e.,g=g
–1
forallgD
2
.Accordingly,
TacticAI’s neural network layers are designed to be D
2
-equivariant.
Notethat,inthecontextof football,D
2
equivariance isnot exact—reflecting
aparticularcornerkickmaynotresultinexactlythesameoutcomes.Among
otherfactors,manycorner-kicktakerstendtohaveapreferredfoot,which
directlyimpactstheprecisionwithwhichtheycancrosstheballintothe
penalty box.However, the dataefficiencybenefits ofconstraininga modelinto
outputting D
2
-equivariant predictions mayoutweigh any inaccuracies in those
predictions—an assumption that turned out to be true
11
.
ConstructingD
2
-equivariantGNNsSinceD
2
isasmall,discretegroup,
the blueprint of group convolutions we described throughout this Chapter per-
fectlyapplies.Firstly,letusassumewehaveatourdisposalagraphneural
network(GNN)layer
12
,G(X),thatcanoperateoverinputnodefeatures,X
(noteweomittheadjacencymatrixhereforbrevity,asitwillalwaysremain
fullyconnected).OurD
2
-equivariantGNNwillcarefullydistributeG across
various views intoacorner kicktacticalsetup,inawaythatpreservestheD
2
symmetry in the outputs (as pointed out by Figure 7.5).
Once wehave ourinput features,XR
22×ばつk
, wefirst "lift"them intothe D
2
space by generatingall fourtransformed views: X
g
=Xρ(g)for gD
2
13
. The
correspondingrepresentationmatricesρ(g)R
k×ばつk
aresimpletoconstructif
ourspatial featuresarezero-centeredaroundthe pitchcenter—we simplyneed
toflipthe signofallcolumnsofXcorrespondingtotheaxes beingreflected.
For example, over vertical flips this amounts tothe following diagonalmatrix:
ρ
ij
=
1i=ji is not a y-axis feature
–1i=ji is a y-axis feature
0i=j
Groups205
Now, we can follow Equation 7.96 to define our group-convolutional layer:
H
g
=
1
4
X
hD
2
G
X
h
X
g
–1
h
whereinwe’vereplacedthe(inner)productwithourGNNlayerGapplied
over concatenatedfeatures.Thislayeryields latentrepresentations H
g
R
22×ばつl
forallviewsgD
2
,andadditionalsuchlayersmaybeeasilystacked.
Finalpredictionsmaybeobtainedthrougheitherframeaveraging(H=
H
e
+H
+H
+H
↔↕
4
)orretrievingH
e
,dependingonwhetheraninvariantor
equivariant prediction is required.
Thisapproachhassuccessfullydeliveredonitspromiseinthelow-data
set-pieceanalyticsdomain:forreceiverandshotprediction,itimprovedthe
baselineGNN’spredictivepowerbyover5%;acomparablejumpwasobtained
byleveragingagraphstructureinthefirstplace(comparedtousingaDeep
Setsmodel).Accordingly,D
2
equivariancebecameoneofthe uniquelynotable
features of TacticAI.
Suggested Further Reading
Wehavepresentedthekey ingredientsthatarerelevantforconstructingarbi-
traryG-equivariantnetworksoverhomogeneousdomainsΩ–includingthe
all-importantG-equivariantnetworksoverGitself!–aswellasseveralrele-
vantexamples. However, for reasonsofbrevity, ourbook deliberatelydoesnot
diveintohowallofthevarious componentsofthesearchitecturesfittogether
more broadly, nor does it detailedly survey how they historically emerged.
TheearliestreferencetogroupCNNs,explicitlyextendingCNNstohan-
dle90-degreerotations,wasprovidedconcurrentlybyT.CohenandWelling
(2016)andDieleman,Fauw,andKavukcuoglu(2016).Makingthesemodels
moreefficientbyleveragingirreduciblerepresentationshasyieldedamore
genericclass ofsteerableCNNs,forwhichwerecommendthepaperbyTaco
S. Cohen and Welling (2017b).
Intheyears followingthesepapers,a widerangeofarchitecturesincorporat-
inggrouporsteerableconvolutionshavebeenproduced,andthesehavebeen
thoroughlysynthesised(withaunifyingtheoryofequivariance over homoge-
neousspaces)byCohen,Geiger,andWeiler(2019).Wewarmlyrecommend
thispaper,notonlyforitssynthesisefforts,butasanexcellentcollectionof
references to other influential works in this space.
Buildingonthesefoundationalideas,subsequentresearchhasfocusedon
scalingthesearchitecturesandextendingthemtocontinuousdomains.Fora
highlypracticalanddefinitive realizationofgeneralsteerableconvolutions—
whichhasbecomeastandardreferenceformodernimplementations—we
206Chapter 7
pointthereadertotheworkofWeilerandCesa(2019b).Furthermore,while
our text remainsfocused onstandarddomains andavoids thebroader topicof
Liegroups,readersinterestedinextendingequivariancetoarbitrarycontinu-
ous groupsoperating onirregular data,such aspoint clouds,will finda critical
milestone in the LieConv architecture proposed by Finzi et al. (2020).
Exercises
1.ConsiderahomogeneousdomainΩ,andasymmetrygroupGactingonit.Given
anyelementuΩ,we canform thestabiliser subgroupG
u
={g| g.u=u},consisting
of those elements of Gthat leave u invariant.
(i)Show that the stabiliser G
u
is indeed a subgroup of G.
(ii)LetG=O(3)bethegroupoforthogonal3×ばつ3matrices(AA
=I,i.e.,thegroup
ofcontinuous3Drotationsandreflections)actingonthesphereΩ=S
2
.Whatisthe
stabiliser of an arbitrary point uΩ (e.g., the north pole)?
(iii)Let G=S
n
be thegroupofpermutationsofnelements,Ω={1,2,...,n}.Whatisthe
stabiliser of an arbitrary point uΩ?
(iv)Letu,vΩbetwoelementsonthedomain.ByhomogeneityofΩ,wecanfinda
g∈Gsuch that g.u=v. Show that, if h∈G
v
, then g
–1
hg∈G
u
.
(v)Fromthepreviouspart,itfollowsthatthereexistsamappingf:G
v
→G
u
definedby
f(h)=g
–1
hg. Provide an inverse of f, i.e., a map f
–1
:G
u
→G
v
.
(vi)Showthatf(hh
)=f(h)f(h
)forallh,h
∈G
v
;i.e.,thatfisagrouphomomor-
phism.This,coupledwiththeexistenceoff
–1
,impliesthatanytwopointsu,vofa
homogeneous space Ωhave isomorphic stabiliser subgroups.
2.Let G=S
n
bethe groupof permutationsacting ona setofn elementsΩ={1,...,n}.
LetS
n–1
bethesubgroupofS
n
thatleavesinvarianttheelement1,i.e.g.1=1for
allgS
n–1
.WewillstudythecosetsgS
n–1
={gh|hS
n–1
}(wheregS
n
)andthe
quotient S
n
/S
n–1
={gS
n–1
| gS
n
} (the set of cosets).
(i)What is the size, N, of S
n
/S
n–1
, i.e., how many cosets are there?
(ii)CharacterisethecosetsgS
n–1
,i.e.,findapropertythatallelementsofagivencoset
have,andthatnoelementoutsideofthecosethas.Inthefollowingparts,wewill
leverage this to identify S
n
/S
n–1
with the set {1,...,N} in the natural way.
(iii)Considerthemapπ:S
n
S
n
/S
n–1
definedbyπ(g)=gS
n–1
.Whatarethefibers,F
u
=
π
–1
(u), where uS
n
/S
n–1
is a coset? (Note: in this context, πis a fiber bundle.)
(iv)Considertheright actionofS
n–1
onS
n
,definedasg.h=gh,for gS
n
,hS
n–1
.Show
that this action preserves fibers, i.e., π(g.h)=π(g) for all gS
n
,hS
n–1
.
(v)Additionally,show thatthisrightactionistransitiveon thefibers,i.e.,forany g,g
π
–1
(u), there exists hS
n–1
such that g.h=g
.
(vi)Additionally, show that thisright action is fixed-pointfree. That is, if g.h=g for some
gS
n
and hS
n–1
, then h=e (the identity).
Groups207
(vii)Asectionofthebundleπ:S
n
S
n
/S
n–1
isamaps:S
n
/S
n–1
S
n
thatsatisfiesπs=
id
S
n
/S
n–1
,i.e.,π(s(u))=uforalluS
n
/S
n–1
.Definesuchasection.(Note:asectionofa
principal bundle is also known as a gauge.)
(viii)Letf:ΩRrepresentascalarfeatureassignedtoeachelement(e.g.anodeina
graph). Apermutation-equivariant linearlayer over thesefeatures isdefined bya kernel
K :Ω×ばつΩRsuchthatK(g.u,g.v)=K(u,v)forallgS
n
.Byconsideringtheaction
of S
n–1
and the results in previous parts, prove Kcan take at most two distinct values.
(ix)LetxR
n
bethevectorofscalarfeaturesforallnelementsofΩ.Writedownthe
most general lineartransformationy=Wxthat is permutation-equivariant, expressing
thematrix Winterms ofthe n×ばつnidentitymatrix Iand theall-ones matrixJ. Basedon
thisformulation,showthatcomputingtheupdatedfeaturesyR
n
requiresonlyO(n)
operations, avoiding the O(n!) complexity of lifting the features into the full group S
n
.
(x)Ifweallowthenodeupdatetobeageneralnon-linearfunctionratherthanalinear
kernel, we specifya permutation-equivariant map F:R
n
R
n
. By leveraging asimilar
argumentasinPart(viii),provethattheoutputfornodeimusttaketheformy
i
=
φ(x
i
,X
i
),whereX
i
istheunorderedmultisetofallothernodefeatures,andφis
somenon-linearfunctionthatispermutation-invariantinitssecondargument.Note
this formulation reflects the Deep Sets architecture (Zaheer et al. 2017).
Notes
1
Technically,weneedthegrouptobelocallycompact,sothatthereexistsa
left-invariantHaarmeasure.Integratingwithrespecttothismeasure,wecan
"shift"theintegrand byanygroupelementandobtainthesameresult,justas
how we have
Z
+
x(u)du=
Z
+
x(uv)du
for functions x:RR.
2
Notethatwhatwedefinehereisnotconvolutionbutcross-correlation,which
istacitlyusedindeeplearningunderthename‘convolution’.Wedoitfor
consistencywiththefollowingdiscussion,sinceinournotation(ρ(g)x)(u)=
x(uv) and (ρ(g
–1
)x)(u)=x(u+v).
3
Theintegrationisdonew.r.t.aninvariantmeasureμonΩ.Incaseμis
discrete, this means summing over Ω.
4
Actually here again, this is cross-correlation.
5
Recall that therepresentation ofGactingon functions definedon Gitself is
called the regular representation of G.
6
The additional properties, e.u=u and g(h.u)=(gh).u, are tacitly assumed.
7
DNAisabiopolymermoleculemadeoffourrepeatingunitscalled
nucleotides(Cytosine,Guanine,Adenine,andThymine),arrangedintotwo
strandscoiledaroundeachotherinadoublehelix,whereeachnucleotide
occurs opposite of the complementary one (base pairs A/T and C/G).
208Chapter 7
8
Also occasionally referred to as "soccer"; a name that we disagree with.
9
LFC’sresearchteamwas,infact,verywellpositionedtoproposeand
improve equivariantarchitectures,asmanyofthem comefrom aphysicsback-
ground,andtheyhavegenerallypioneereddata-drivenapproachestofootball
analyticsovertheyears;seeGraham(2024)forabeautiful,football-centric
account of these successes.
10
The dihedralgroup D
2
is alsoisomorphic tothe Kleinfour-group, K
4
, first
describedbyFelixKleinin1884.Abstractly,K
4
={e,a,b,c}isagroupof
fourelementswhereeachelementisself-inverse(gg=eforallgK
4
)and
composinganytwonon-identitiesyieldsthethirdelement(ab=ba=c,ac=
ca=b,bc=cb=a).K
4
hasdiverseapplications, fromalgebra(whereitcanbe
used toexplaintheexistenceof formulaefor solvingquarticequations), allthe
way to music composition (being related to dodecaphony).
11
Generally,about30%oftheplayersinthePremierLeagueareleft-footed
or both-footed, which gives a sufficient amount of players who can effectively
attackaballinamirroredcapacity.Asanexample,mostteamshavealeft-
footedleft-backandaright-footedright-backwhotakecorners,anditwould
beraretofindateamthatdoesn’thavebothaleft-footedandright-footed
corner taker. Interms of attackingthe corners, headedshots are very common
incornerkicksituations,sofootednesstendstomattersignificantlylessin
these situations.Also, manyshots takenduring cornersare morereactive from
close range, where foot preference matters less.
12
Specifically,TacticAIusestheGATv2(Brody,Alon,andYahav2022)as
the implementation of G.
13
Note that,unlikemanypreviousexamples,herethe representationρ(g) acts
viaright-multiplication,asitactsonthenodefeaturesratherthanthenodes
themselves.

AltStyle によって変換されたページ (->オリジナル) /