Masked inference and similarity increase across model hierarchy.

<p><b>top</b>. We visualize the latent layer activity for a masked input (top left) across all latent layers for PC and MCPC, following the method outlined in Section 4.2.2 of the manuscript. The MCPC model identifies different potential interpretations for a given masked input across its latent layers, whereas the PC model infers only one possible interpretation. Additionally, we visualize the reconstructed images by the MCPC model when the latent layers represent two different possible interpretations. The reconstructed digit when the MCPC model infers a “4” resembles the digit four, while the reconstructed digit when the model infers a “9” resembles the digit nine. <b>bottom</b>. We repeat the analysis to assess the similarity between spontaneous activity and average evoked activity across all latent layers, following the method described in Section 4.3.2. The KL divergence between spontaneous activity and average evoked activity for natural stimuli is lower compared to noise images and image gratings for an MNIST-trained MCPC model across all latent layers. However, this difference is only statistically significant in the first latent layer <i>x</i><sub>1</sub> and not in the higher latent layers.</p> <p>(PDF)</p>