Moving Past the Minimum Information Partition: How To Quickly and Accurately Calculate Integrated Information

Daniel Toker,Friedrich Sommer

DOI: https://doi.org/10.48550/arXiv.1605.01096

2016-05-04

Abstract:An outstanding challenge with the Integrated Information Theory of Consciousness (IIT) is to find a way of rapidly and accurately calculating integrated information from neural data. A number of measures of integrated information based on time series data have been proposed, but most measures require finding the Minimum Information Partition of a network, which is computationally expensive and not practical for real brain data. Here, we introduce a novel partition, the Maximum Modularity Partition, across which to quickly calculate integrated information. We also introduce a novel detection task on simulated data to evaluate the performance of integrated information measures across different partitions. We show that integrated information can be reliably and quickly calculated across the Maximum Modularity Partition, as well as the previously proposed atomic partition, even in relatively large networks, which constitutes an advance in researchers' ability to empirically test the predictions of IIT in real brain data.

Neurons and Cognition

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: how to quickly and accurately calculate integrated information from neural data. The Integrated Information Theory (IIT) believes that the level of consciousness in the brain can be measured by the amount of its integrated information. However, existing calculation methods either require systematic perturbation of each node in the network, which is impossible to achieve in actual brain data; or they need to find the "Minimum Information Partition" (MIP) of the network, and this process has extremely high computational costs and is not suitable for real - brain data. Therefore, the paper proposes a new partitioning method - Maximum Modularity Partition (MMP), and evaluates the performance of different partitioning methods in calculating integrated information through simulation data detection tasks.

### Main contributions:
1. **Propose Maximum Modularity Partition (MMP)**: This is a new network partitioning method for quickly calculating integrated information. Compared with MIP, MMP is not only fast in calculation but also more in line with the actual information flow pattern in the brain.
2. **Introduce a new detection task**: Generate time series through simulation data and test the performance of different integrated information metrics when detecting network cuts.
3. **Compare the performance of different partitioning methods**: Study the accuracy and calculation speed of MMP, atomic partition and Minimum Information Bipartition (MIB) when calculating integrated information.

### Method overview:
- **Integrated information metrics**: The paper tests three integrated information metric methods based on time - series data: Causal Density (CD), \(\Phi^*\) and \(\Phi_{AR}\).
- **Maximum Modularity Partition (MMP)**: Use the Louvain algorithm to find the maximum modularity partition of the network, and the time complexity of this method is \(O(n \log n)\).
- **Detection task**: Generate multivariate time - series data and test the performance of different integrated information metrics when detecting network cuts.

### Results:
- **Detection performance**: CD performs best in the detection task, followed by \(\Phi_{AR}^{MMP}\) and \(\Phi_{AR}^{atom}\). MMP and atomic partition are significantly faster in calculation than MIB.
- **Calculation speed**: The calculation speed of MMP and atomic partition is very fast, and the calculation can be completed within a few seconds even in a relatively large network, while the calculation time of CD increases exponentially with the network scale.
- **Correlation analysis**: The results of different integrated information metrics on different networks are not completely consistent, but the results of MMP and atomic partition are relatively close.

### Conclusion:
The Maximum Modularity Partition (MMP) method proposed in the paper is not only fast in calculation but also performs excellently in detecting network integrated information, providing researchers with a practical tool to test the predictions of IIT in real - brain data.