AI-driven modelling for cortex-wide neuromodulated learning

Both animals and humans can learn cognitive tasks from sensory inputs. However, existing computational frameworks are extremely slow at learning such tasks. Therefore, a new generation of computational frameworks capable of rapid learning of cognitive tasks is urgently needed.
Here we propose that neuromodulation together with specific excitatory-inhibitory circuits enable efficient learning across multiple brain areas. Building on our expertise in developing biologically plausible AI-driven computational models of learning we will develop an adaptive synapse-to-behaviour model of neuromodulated cortex-wide learning. First, we will show that a precise neuromodulatory control of cell-types results in rapid brain-wide learning of cognitive tasks, but also in neural networks that are more robust to perturbations. Next, the predictions generated by the model will be tested using existing experimental data. We will test the model at the cellular, systems and behavioural level. Overall, the proposed integrative AI-driven computational framework, will be critical for our understanding of cortex-wide learning of cognitive tasks in both health and disease.

Learning of cognitive tasks depends on behavioural feedback. In the brain this feedback is encoded by diffuse neuromodulators, such as those released by the Cholinergic system. These neuromodulators are believed to trigger learning by controlling circuits throughout the cortex. However, existing computational and conceptual frameworks of cortex-wide learning exhibit very slow learning, in sharp contrast with animal and human learning. This raises a fundamental question: how can the cortex learn rapidly using diffuse neuromodulation?

Here we propose that Cholinergic neuromodulation relies on specific excitatory-inhibitory circuits to assign credit efficiently to synapses throughout the cortex. Building on our expertise in developing AI-driven computational models, systems, cellular and synaptic neuroscience of cortical learning we will develop an adaptive synapse-to-behaviour model of neuromodulated cortex-wide learning. First, we will show that a precise control of cell-types results in rapid and more robust brain-wide learning of cognitive tasks, but also in task-encoding that are more robust to perturbations. Next, to demonstrate the plausibility of the model, it will be contrasted with recently acquired experimental data. We will first perform tests at the cellular level, by studying how feedforward and feedback cortical pathways should be controlled by the Cholinergic system. In addition, we will test our model at the systems level by comparing the control of inhibitory cell-types predicted by our model with imaging of the same cell-types during goal-driven learning in the cortex.
Overall, the proposed integrative AI-driven computational framework, will be critical for our understanding of cortex-wide learning of cognitive tasks in both health and disease.