Appendix Table A1.

Network capacity for model comparison.

Architecture	Config.	$w$	Capacity	Dev. (%)
Encoder	$C$ = $2^0$	72	252 k	$+0.8$
Encoder	$C$ = $2^1$	45	250 k	$0.0$
Encoder	$C$ = $2^2$	27	247 k	$+1.2$
Encoder	$C$ = $2^3$	16	255 k	$+2.0$
Encoder	$C$ = $2^4$	9	260 k	$+4.0$
Decomposer	$n$ = 0	63	253 k	$+1.2$
Decomposer	$n$ = 1	59	254 k	$+1.6$
Decomposer	$n$ = 2	55	248 k	$-0.8$
Decomposer	$n$ = 3	52	246 k	$-1.6$
Detector	$k$ = $2^3$	62	248 k	$-0.8$
Detector	$k$ = $2^4$	62	252 k	$+0.8$
Detector	$k$ = $2^5$	61	253 k	$+1.2$
Detector	$k$ = $2^6$	58	246 k	$-1.6$
Encoder	$C$ = $2^0$	110	498 k	$-0.4$
Decomposer	$n$ = 0	89	504 k	$+0.8$
Detector	$k$ = $2^5$	87	503 k	$+0.6$

Each model configuration uses a different network width $w$ held constant across its convolutional layers to reach a target capacity of 250 k or 500 k parameters, up to a small deviation.