. 2022 Apr 19;13:2064. doi: 10.1038/s41467-022-29632-7

Table 3.

Latent sparsity measurements.

	GNCN-t1-Σ/Friston			GNCN-t2-LΣ			GNCN-PDH
	$ρ_{a}^{1}$	$ρ_{a}^{2}$	$ρ_{a}^{3}$	$ρ_{a}^{1}$	$ρ_{a}^{2}$	$ρ_{a}^{3}$	$ρ_{a}^{1}$	$ρ_{a}^{2}$	$ρ_{a}^{3}$
MNIST	0.74	0.60	0.63	0.21	0.18	0.16	0.22	0.21	0.19
KMNIST	0.35	0.50	0.67	0.24	0.21	0.18	0.26	0.24	0.21
FMNIST	0.41	0.44	0.68	0.21	0.20	0.18	0.24	0.22	0.20
CalTech	0.50	0.68	0.69	0.21	0.19	0.17	0.25	0.23	0.20

Sparsity levels in the GNCN-t1-Σ/Friston⁵ versus the GNCN-t2-LΣ and GNCN-PDH. $ρ_{a}^{ℓ}$ indicates sparsity for layer ℓ in each model on each of the four databases— $ρ_{a}^{1}$ is the measured sparsity level for the first layer of latent neural activities, $ρ_{a}^{2}$ is the sparsity level of the second layer, and $ρ_{a}^{3}$ is the sparsity level for the third layer (sparsity analysis was performed using all data samples in the training set.).