Artificial Life
November 5, 2024
Carlotta Langer, Nihat Ay
Using simulated agents, the authors model how an embodied agent and its environment influence each other via a sensorimotor loop. Information-theoretic measures quantify information flows, including morphological computation (interaction between body and environment) and controller complexity, which relates to integrated information theory of consciousness. Prior work found that a well-adapted morphology reduces needed controller complexity. Here, the authors observe that agents must first understand relevant environmental dynamics to interact effectively, so increased controller complexity can improve body-environment interaction.
arXiv Preprint Archive
August 2, 2021
Carlotta Langer, Nihat Ay
Integrated Information Theory offers a quantitative method for studying consciousness and can be applied to neural networks. This paper combines methods to examine information flows among the body, brain, and environment of an embodied agent. Using a simple experimental setup, the authors calculate optimal policies for goal-directed behavior via a "planning as inference" method, then compute morphological computation and integrated information relative to those policies. Comparing these measures under changing morphological conditions reveals an antagonistic relationship: the more morphological computation is involved, the less information integration within the brain is required. The authors argue that measuring information flow to and from the brain is necessary to determine the brain's influence on behavior.
Frontiers in Psychology
January 1, 2021
Carlotta Langer, Nihat Ay
The Integrated Information Theory offers a quantitative way to study consciousness and can be applied to neural networks. An embodied agent controlled by such a network interacts with its environment, involving morphological computation in goal-directed action and integrated information within the controller, the agent's brain. This article combines methods to examine information flows among and within the body, brain, and environment of an agent. In a simple experimental setup, the optimal policy for goal-directed behavior is calculated using the planning as inference method, where the information-geometric em-algorithm optimizes the likelihood of the goal.
Entropy (Basel, Switzerland)
September 30, 2020
Carlotta Langer, Nihat Ay
A new measure of integrated information, called ΦCII (Causal Information Integration), is proposed as an alternative to existing measures that quantify the strength of causal connections between neurons in the context of Integrated Information Theory of consciousness. Unlike the candidate measure ΦCIS, which lacks a graphical representation and is difficult to analyze, ΦCII satisfies all desirable properties and can be calculated using an iterative information geometric algorithm (the em-algorithm). This allows comparison with existing integrated information measures.
arXiv Preprint Archive
August 26, 2020
Carlotta Langer, Nihat Ay
A new measure of integrated information, Φ_{CII}, is proposed for quantifying causal connections between neurons within the Integrated Information Theory of consciousness. Unlike a prior candidate, Φ_{CIS}, which satisfies all theoretically desirable properties but lacks an intuitive graphical representation, Φ_{CII} models a common exterior influence via a latent variable and also satisfies all required conditions. It can be computed with an iterative information-geometric algorithm (the em-algorithm), enabling comparison with existing integrated information measures.