Causal
representation learning (CRL) aims at learning causal factors and their ca
usal relations from high-dimensional observations\, e.g. images. In genera
l\, this is an ill-posed problem\, but under certain assumptions or with t
he help of additional information or interventions\, we are able to guaran
tee that the representations we learn are corresponding to some true under
lying causal factors up to some equivalence class.
\nIn this talk I w
ill first present CITRIS (https://proceedings.ml
r.press/v162/lippe22a/lippe22a.pdf)\, a variational autoencoder framew
ork for causal representation learning from temporal sequences of images\,
in systems in which we can perform interventions. CITRIS exploits tempora
lity and observing intervention targets to identify scalar and multidimens
ional causal factors\, such as 3D rotation angles. In experiments on 3D re
ndered image sequences\, CITRIS outperforms previous methods on recovering
the underlying causal variables. Moreover\, using pretrained autoencoders
\, CITRIS can even generalize to unseen instantiations of causal factors.<
br />\n
\nWhile CRL is an exciting and promising new field of researc
h\, the assumptions required by CITRIS and other current CRL methods can b
e difficult to satisfy in many settings. Moreover\, in many practical case
s learning representations that are not guaranteed to be fully causal\, bu
t exploit some ideas from causality\, can still be extremely useful. As ex
amples\, I will describe some of our work on exploiting these "causality-i
nspired" representations for adapting policies across domains in RL (https://openreview.net/forum?id=8H5bpVwvt5) and to nonstationar
y environments (https://openreview.net/forum?id=VQ9fogN1q6e)\, and how learning a factored graphical representations (even if not n
ecessarily causal) can be beneficial in these (and possibly other) setting
s.