Caffeine
What I want to do
Euclidean group
SE(3) = rotation (SO(3)) + translations
(∘,G) is a group
(V,D) is a representation of the group G
irreducible representations of G
Direct sum
Tensor product
⊕
=
⊗
=
irreducible representations of G
⊗
=
Caffeine
What I want to do
input data
Fa
Fb
Equivariance
f(D(R)x)=D′(R)f(x)
Fa′
xba
Fb
Fa′
xba
Fb
Fa′=b∑∥xba∥Fb
Fa′
xba
Fb
Fa′=b∑∥xba∥Fb
Fa′=b∑xba∧Fb
Fa′
xba
Fb
Fa′=b∑xba(xba⋅Fb)
Fa′=b∑∥xba∥Fb
Fa′=b∑xba∧Fb
Fa′=b∑∥xba∥Fb
Fa′=b∑xba∧Fb
Fa′=b∑xba(xba⋅Fb)
Can we formalize this a bit?
Fa′
xba
Fb
irreducible representations of SO(3)
DL(g)
for L=0,1,2,3,…
dim(DL)=2L+1
scalars (color, temperature, time left, ...)
vectors (velocity, gravity, dipole moment, ...)
traceless symmetric tensor (moment of inertia, stress tensor, ...)
3⊗3=1⊕3⊕5
D1⊗D1=D0⊕D1⊕D2
Spherical harmonics
YL(Rx)=DL(R)YL(x)
0
+1
-1
Y0
Y1
Y2
YL:s2⟶R2L+1
Fa′=b∑R(∥xab∥,ω)YL(∥xab∥xab)⊗Fb
learned function
spherical harmonics
tensor product
linear
equivariant
1-neighbor
our method
Relevance of Rotationally Equivariant Convolutions for Predicting Molecular Properties, K. Miller et al.
ongoing development
many possible architectures to explore
thanks :)
http://e3nn.org