Skip to content Skip to navigation

Mind, Brain, Computation and Technology graduate training seminar - Dan Birman & Ian Eisenberg

June 3, 2019 - 5:10pm
Sloan Hall, Math Building 380, Room 380-C

Flexible readout as a selection mechanism of human sensory representations

Dan Birman
Mind, Brain, Computation and Technology graduate trainee, Stanford University



Daily life requires that people flexibly attend to different aspects of the environment when their goals require it. This can be operationalized by asking human observers to report about one feature of a visual stimulus while ignoring others. Such context-dependent behavior could be supported by many different cortical mechanisms. Implementations could either change sensory representations or change the "readout" according to task demands. In this talk I explore how readout occurs when observers are asked to report about the visibility of motion. We developed a linking model to connect measurements of behavior and BOLD signal. The model shows that although changes in sensory representation occur as observers shift from reporting one feature to another, these changes are insufficient to explain changes in behavior. Instead, we propose that changes in readout allow observers to shift their behavior according to task demands.

Curriculum Vitae

Related papers

[1] Birman, D. and Gardner, J.L. (2018). A quantitative framework for motion visibility in human cortex. J Neurophysiol 120: 1824 –1839. 


Ian Eisenberg, Mind, Brain, Computation and Technology Trainee

Uncovering mental structure through data-driven ontology discovery

Ian Eisenberg
Mind, Brain, Computation and Technology graduate trainee, Stanford University


Cognitive neuroscience has linked neural activity to a wealth of cognitive processes, yet struggles to produce a cumulative account of neural function. This slow progress has many causes, but is partially explained by the lack of systematic ontologies describing brain structure and mental function. While integrative brain atlases have been steadily improving, commensurate efforts to improve cognitive ontologies have been limited. We address this by developing a data-driven cognitive ontology derived from individual differences across a broad range of behavioral tasks, self-report surveys, and real-world outcomes. 522 participants completed 62 different measures on Mechanical Turk related to decision-making, working memory, cognitive control, impulsivity, and personality, amongst other psychological constructs. Interestingly, though subsets of the tasks and surveys putatively reflect similar constructs, we find that they bifurcate in the ontology. Using exploratory factor analysis, we identify two low-dimensional cognitive spaces that separately capture behavioral tasks and surveys. Hierarchical clustering within these spaces identify sensible clusters which capture psychological "kinds", related to, but separate, from the dimensions identified with factor analysis. Overall, this structure discovery reveals a simpler cognitive ontology than typically employed in the psychological sciences. As real-world relevance is an essential feature of theoretical constructs, we also evaluated whether tasks and surveys can predict real-world outcomes. We reduced the self-reported real-world outcomes to 8 "target" factors (e.g. mental health, binge drinking), computed individual factor scores, and assessed predictive ability using cross-validated ridge regression. While surveys performed moderately well, tasks showed almost no predictive ability.
Cognitive ontologies describe the psychological constructs through which most human neuroscience is understood. We demonstrate that data-driven structure discovery techniques can profitably improve these ontologies.

Curriculum Vitae

Related papers (coming soon)