Presenter
Nathan Helm-Burger
Nathan studied neuroscience in graduate school, then spent five years working in industry as a data scientist and machine learning engineer. He has spent the last four years studying AI alignment and safety. His research topics have been: AI capability forecasting, AI Biorisk Evaluation, Corrigibility, and how we can learn from neuroscience to improve AI interpretability.
Summary:
Recent data from the Human Connectome Project has revealed surprisingly narrow informational bottlenecks. The emerging picture of a modular structure with a mix of highly dynamic boundaries and static bottlenecks. This compute graph suggests a new architecture for machine learning models with greatly improved functional localization. If the new architecture proves competitive with existing models, the functional localization would greatly improve interpretability and steerability of the resulting models.
