Computational Account of Spontaneous Activity as a Signature of Predictive Coding
Fig 8
Characteristics of the optimal network.
A: Frequency of the Up states as a function of the linear cost. Frequency of Up states decreases with increasing linear cost. At the optimal cost, indicated by an arrow, the frequency of Up states is close to zero, indicating that in the optimally efficient network, Up states are at the point of vanishing. b) Optimal linear cost in function of the strength of the noise. Optimal linear cost is modulated by the strength of the noise (i.e. the σ of the noise process) in a non-monotonous fashion. For small noise levels, the optimal cost is decreasing, reaches a minimum and increases thereafter. This is true for the optimal cost estimated from the Total error (black trace), as well as for the optimal cost estimated from the coding error only (red trace). c) Frequency of Up states for the networks with optimal costs in function of the strength of the noise. At the optimum, frequency of Up states is always close to zero, irrespective of the level of the noise. This is true for both active (full line) and quiescent state activity (dashed line). d) Optimal linear cost does not depend on the strength of the input. For a reasonable range of input strengths, optimal cost stays constant. All other parameters are in the Table 2 in S1 Table.