Periodic Reporting for period 3 - See-ACC (Cracking the Anterior Cingulate Code: Toward a Unified Theory of ACC Function)
Reporting period: 2022-07-01 to 2023-12-31
Anterior cingulate cortex is one of the largest riddles in cognitive neuroscience and presents a major challenge to mental health research. ACC dysfunction contributes to a wide spectrum of psychiatric and neurological disorders but no one knows what it actually does. Although more than a thousand papers are published about it each year, attempts to identify its function have been confounded by the fact that a multiplicity of tasks and events activate ACC, as if it were involved in everything.
Recently, I proposed a theory that reconciles many of the complexities surrounding ACC. This holds that ACC selects and motivates high-level, temporally extended behaviors according to principles of hierarchical reinforcement learning. For example, on this view ACC would be responsible for initiating and sustaining a run up a steep mountain. I instantiated this theory in two computational models that make explicit the theory's assumptions, while yielding testable predictions. In this project my team is integrating the two computational models into a unified, biologically-realistic model of ACC function. We are systematically testing the model in a series of experiments involving functional magnetic resonance imaging, electroencephalography and psychopharmacology, in both healthy human subjects and patients.
The establishment of a complete, formal account of ACC will fill an important gap in the cognitive neuroscience of cognitive control and decision making, strongly impact clinical practice, and be important for artificial intelligence and robotics research, which draws inspiration from brain-based mechanisms for cognitive control. The computational modelling work will also link high level, abstract processes associated with hierarchical reinforcement learning with low level cellular mechanisms, enabling the theory to be tested in animal models.
Recently, I proposed a theory that reconciles many of the complexities surrounding ACC. This holds that ACC selects and motivates high-level, temporally extended behaviors according to principles of hierarchical reinforcement learning. For example, on this view ACC would be responsible for initiating and sustaining a run up a steep mountain. I instantiated this theory in two computational models that make explicit the theory's assumptions, while yielding testable predictions. In this project my team is integrating the two computational models into a unified, biologically-realistic model of ACC function. We are systematically testing the model in a series of experiments involving functional magnetic resonance imaging, electroencephalography and psychopharmacology, in both healthy human subjects and patients.
The establishment of a complete, formal account of ACC will fill an important gap in the cognitive neuroscience of cognitive control and decision making, strongly impact clinical practice, and be important for artificial intelligence and robotics research, which draws inspiration from brain-based mechanisms for cognitive control. The computational modelling work will also link high level, abstract processes associated with hierarchical reinforcement learning with low level cellular mechanisms, enabling the theory to be tested in animal models.
The project consists of the following 4 work modules:
Module 1: Development of a Unified Module of ACC Function.
We have created a hierarchically-organized recurrent neural network model of ACC, which can implement complex, goal-directed multi-step action sequences. Crucially, the model predicts how hierarchical action sequences are represented at different levels of abstraction along the expanse of ACC. These results are being prepared for publication (Colin, Ikink & Holroyd, in preparation). We are also applying the model to make predictions for several functional magnetic resonance imaging (fMRI) experiments that we are conducting concurrently.
Module 2 and Module3: Hierarchical Represenations in ACC, Control Over Action Sequences These modules consists of fMRI experiments that test the computational models developed in Module 1. We have collected data for four experiments (including the preliminary data that were submitted with the grant proposal) and are now starting data collection for the fifth. Crucially, application of the model developed in Module 1 to the data collected for the first experiment in this series (“Task Representations in ACC”) has confirmed the predicted anterior-posterior hierarchy along ACC, with anterior ACC representing high-level goals (e.g. making tea) and posterior ACC representing low-level actions (e.g. stirring) (see attached figure). These results are being prepared for publication (Colin, Ikink & Holroyd, in preparation).
Module 4: Causal Evidence for Task Regulation by ACC.
We have begun data collection for 2 experiments in this module. First, about 20 patients with stroke damage to frontal cortex have participated in tasks that examine the effect of this damage on 1) the learning and production of hierarchical action sequence, and 2) the generation of a reward signal thought to be produced in ACC. Second, about 8 patients with epilepsy have participated in a study in which we record reward-related brainwave activity using electrodes placed inside the brain. Because these two experiments require large numbers of participants, data collection will not be completed for several years yet.
Other: We have further developed a set of behavioral and EEG experiments that examine how ACC learns hierarchical representations. Data collection for these experiments has recently begun. Lastly, we have written a review article about the computational mechanisms of ACC, which lays out the ideas of this grant (Holroyd & Verguts, 2021).
Module 1: Development of a Unified Module of ACC Function.
We have created a hierarchically-organized recurrent neural network model of ACC, which can implement complex, goal-directed multi-step action sequences. Crucially, the model predicts how hierarchical action sequences are represented at different levels of abstraction along the expanse of ACC. These results are being prepared for publication (Colin, Ikink & Holroyd, in preparation). We are also applying the model to make predictions for several functional magnetic resonance imaging (fMRI) experiments that we are conducting concurrently.
Module 2 and Module3: Hierarchical Represenations in ACC, Control Over Action Sequences These modules consists of fMRI experiments that test the computational models developed in Module 1. We have collected data for four experiments (including the preliminary data that were submitted with the grant proposal) and are now starting data collection for the fifth. Crucially, application of the model developed in Module 1 to the data collected for the first experiment in this series (“Task Representations in ACC”) has confirmed the predicted anterior-posterior hierarchy along ACC, with anterior ACC representing high-level goals (e.g. making tea) and posterior ACC representing low-level actions (e.g. stirring) (see attached figure). These results are being prepared for publication (Colin, Ikink & Holroyd, in preparation).
Module 4: Causal Evidence for Task Regulation by ACC.
We have begun data collection for 2 experiments in this module. First, about 20 patients with stroke damage to frontal cortex have participated in tasks that examine the effect of this damage on 1) the learning and production of hierarchical action sequence, and 2) the generation of a reward signal thought to be produced in ACC. Second, about 8 patients with epilepsy have participated in a study in which we record reward-related brainwave activity using electrodes placed inside the brain. Because these two experiments require large numbers of participants, data collection will not be completed for several years yet.
Other: We have further developed a set of behavioral and EEG experiments that examine how ACC learns hierarchical representations. Data collection for these experiments has recently begun. Lastly, we have written a review article about the computational mechanisms of ACC, which lays out the ideas of this grant (Holroyd & Verguts, 2021).
Because data analysis for most of our experiments has only just begun, it is premature to evaluate the current impact of this project. Nevertheless, we continue to expect that establishing a complete, formal account of ACC will fill an important gap in the cognitive neuroscience of cognitive control and decision making, strongly impact clinical practice, and be important for artificial intelligence and robotics research, which draws inspiration from brain-based mechanisms for cognitive control. The computational modelling work will also link high level, abstract processes associated with hierarchical reinforcement learning with low level cellular mechanisms, enabling the theory to be tested in animal models.