Mutual information and redundancy in spontaneous communication between cortical neurons

Abstract

An important question in neural information processing is how neurons cooperate to transmit information. To study this question, we resort to the concept of redundancy in the information transmitted by a group of neurons and, at the same time, we introduce a novel concept for measuring cooperation between pairs of neurons called relative mutual information (RMI). Specifically, we studied these two parameters for spike trains generated by neighboring neurons from the primary visual cortex in the awake, freely moving rat. The spike trains studied here were spontaneously generated in the cortical network, in the absence of visual stimulation. Under these conditions, our analysis revealed that while the value of RMI oscillated slightly around an average value, the redundancy exhibited a behavior character- ized by a higher variability. We conjecture that this combination of approximately constant RMI and greater variable redundancy makes information transmission more resistant to noise disturbances. Furthermore, the redundancy values suggest that neurons can cooperate in a flexible way during information transmission. This mostly occurs via a leading neuron with higher transmission rate or, less frequently, through the information rate of the whole group being higher than the sum of the individual information rates-in other words in a synergetic manner. The proposed method applies not only to the stationary, but also to locally stationary neural signals.

Neuronal Assembly Dynamics in Supervised and Unsupervised Learning Scenarios

“…Also, the results showed that, in our model, most of the neuronal architectures were highly redundant, most of the neurons in the higher performance configurations presented independent activity, and that increasing the number of neurons in a network with fixed number of assemblies increased the redundancy. All these findings resemble the results obtained in real cortical experiments, therefore a few remarks should be made: first, real neuronal ensembles are highly redundant, and that can be associated with resistance to error and natural mechanisms of probability distribution estimation (Barlow, 2001;Szczepanski et al, 2011); second, neuronal independence (as observed in our results) can be linked to code efficiency because the information capacity of individual neurons is not compromised by redundant scenarios (Schneidman et al, 2003); and third, there is still a lack of studies comparing the information flow dynamics due to neuronal interactions and due to single neurons alone -attesting the time-scales of the interactions as well as spurious effects such as averaging is still work in progress (Reich et al, 2001;Narayanan et al, 2005).…”

“…The intuition that noisy scenarios are better tackled with redundant architectures is justified -there are works showing that cortical circuits, which operate in an intrinsically noisy environment, are highly redundant (Narayanan et al, 2005;Szczepanski et al, 2011). However, there is criticism regarding the interpretation of information theoretical measurements such as redundancy (Schneidman et al, 2003;Latham & Nirenberg, 2005), as well as findings showing predominantly synergistic or independent activity in neuronal circuits, instead of redundancy, depending on factors such as which area and which neurons are recorded or which kind of task is performed (Reich et al, 2001).…”

“…However, there is criticism regarding the interpretation of information theoretical measurements such as redundancy (Schneidman et al, 2003;Latham & Nirenberg, 2005), as well as findings showing predominantly synergistic or independent activity in neuronal circuits, instead of redundancy, depending on factors such as which area and which neurons are recorded or which kind of task is performed (Reich et al, 2001). Addi- tionally, as shown by Szczepanski et al (2011), neurons can dynamically switch their interactions during the execution of the task, thus synergetic, independent, or redundant activity may be masked by averaging processes. Finally, redundancy and synergy are found to be largely influenced by the network architecture and the decoding unit used (Schneidman et al, 2003).…”

Synchronisation effects on the behavioural performance and information dynamics of a simulated minimally cognitive robotic agent

“…We are mainly interested in studying how ET E between these variables vary as the task progresses, given that the agent continually engages with the object during the discrimination and orientation processes. For this purpose, a sliding window technique is used (Staniek & Lehnertz, 2008;Szczepanski et al, 2011), with a window size of 200 data points. Therefore, at every time step of the task, the ET E is estimated (according to Equations 5 and 6) considering the time series contained in that window.…”

Estimating Temporal Causal Interaction between Spike Trains with Permutation and Transfer Entropy

“…The question of how neurons interact has gained considerable attention over the years [1], [2]. Traditional methods emphasize the strength of pairwise connections, i.e.…”