Controlled analysis of neurocontrollers with informational lesioning

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Systems biology

Background:

• Understanding neural information processing requires identifying the roles of individual network elements.
• Functional Contribution Analysis (FCA) is used to study neurocontrollers in evolutionary autonomous agents (EAAs).
• Previous FCA studies showed dependence of contribution values (CVs) and prediction error on lesioning methods.

Purpose of the Study:

• Introduce a novel informational lesioning method (ILM) for FCA.
• Investigate the impact of ILM on FCA prediction accuracy and CVs.
• Explore how varying lesioning levels reveal element importance in neurocontrollers.

Main Methods:

• Developed and applied an informational lesioning method (ILM) viewing lesions as noisy channels.
• Controlled lesioning by varying lesioning levels from large to small magnitudes.
• Integrated ILM within the Functional Contribution Analysis (FCA) framework using data from multiple lesion experiments.

Main Results:

• Lower lesioning levels with ILM lead to more accurate FCA predictions.
• Minute ILM lesioning levels effectively uncover long-term effects of network elements.
• As lesioning levels decrease, CVs converge to limit values, indicating element importance in intact networks.

Conclusions:

• The ILM enhances FCA by providing more accurate predictions of EAAs' performance under lesions.
• ILM facilitates a deeper understanding of individual element contributions and their long-term impact.
• This method refines the analysis of neural information processing and neurocontroller function.