 
Summary: Supporting Text
Methods
Bayesian chunk learner (BCL)
Prior distributions An inventory I was defined as the set of chunks, and for each chunk the shapes it
influenced (referred to as links). The prior probability of an inventory (used for computing the posterior over
inventories in Eq. 7, and also used in Eq. 8) depended on the number of chunks H and the total number of
links L:
P(I) = P(H) P(LH) (S1)
The prior distribution of the parameters of an inventory I (used in Eqs. 6 and 8) included appearance
probability parameters W = wij, wxi , wyj , and spatial position parameters C = cij, cxi , cyj , ij, xi , yj :
P(II) = P(WI) P(CI) (S2)
The prior distribution of appearance probability parameters W had a simple factorized form
P(WI) =
i
P(wxi )
j
P wyj
j,iparI(j)
P(wij) (S3)
where parI(j) is the set of chunks that influence (have links to) shape j according to inventory I.
