We conducted computational experiments estimating Tononi’s Phi coefficient to measure the “integrated information” within the OpenCog cognitive architecture on two types of tasks: Reading (i.e. parsing and semantically analyzing) short documents and guiding the Sophia humanoid robot in carrying out a dialog-based interaction. The data used to calculate Phi comprises time series of STI (Short Term Importance) values corresponding to Atoms (nodes and links) in OpenCog’s Attentional Focus. To make these computations feasible, we preprocessed our data using Independent Component Analysis. We fed the reduced set of time series into software that applies known methods for approximating Phi. Qualitatively (and preliminarily), comparison of the variation of Phi with cognitive system behavior over time reveals sensible patterns.
Matthew Iklé, Ben Goertzel, Misgana Bayetta, George Sellman, Comfort Cover, Jennifer Allgeier, Robert Smith, Morris Sowards, Dylan Schuldberg, Man Hin Leung, Amen Belayneh, Gina Smith, and David Hanson.