Fig 1.
The infographic represents a simplified application of the Harmonium Model, following the example in the text. A tennis player looking at their opponent’s serve is faced with a very complex perception. The phase space of meaning is a multidimensional space where any ‘object’ of perception (in this example, the ball) is represented in a broad range of dimensions. Not all these dimensions are needed to achieve a precise forecast of the ball’s trajectory, and thus inform future action (where to move to return the serve). The harmonium mechanism operates a dimensionality reduction in order to focus cognitive resources on the dimensions crucial to the task at hand.
Fig 2.
Architecture of the deep learning model.
The input image was vectorised and presented through an input layer (900 neurons). Two hidden layers of 400 and 800 neurons, respectively, were then used to learn a hierarchical generative model in a completely unsupervised fashion (undirected arrows). The behavioural task is finally simulated by stacking a read-out classifier on top of the network (directed arrows).
Fig 3.
Images from the low entropy dataset (TC-LE) contain letters printed using a limited number of variability factors, while images from the high entropy dataset (TC-HE) were created using a greater variety of fonts and styles.
Fig 4.
Schematic representation of the analytic procedures employed in the study.
The diagram exemplifies the procedures of one TC-LE and one TC-HE network on the ‘A’ letter set of inputs. The same steps were repeated for all other letters and networks. Note that the figure’s fonts are exaggeratedly different in order to better convey the study design. For an example of the actual fonts employed, refer to Fig 3. PCs = principal components.
Fig 5.
The panels explain the decisional process employed to choose the number of components in WGHT_PD and WGHT_SD, based on the data of Study 1. Specifically, the left panel shows the distribution (median and 95% range) of the cumulative explained variance of the first five PCs across all PCAs. The right panel is a scree-plot representing PCs 5 to 9 focusing on the eigenvalue = 1 cut-off.
Fig 6.
Comparison between TC-LE and TC-CE PCA distributions.
The lines represent the cumulative percentage of explained variance for the first 100 components of TC-LE and TC-CE networks. A lower explained variance per factor implies a higher entropy.