neocortex; neuromorphic; learning rule; mushroom body; dendrites; error propagation; microcircuit
Marti Mengual Ulisses, Wybo Willem A.M., Spierenburg Lotte J.E., Santello Mirko, Senn Walter, Nevian Thomas (2020), Efficient Low-Pass Dendro-Somatic Coupling in the Apical Dendrite of Layer 5 Pyramidal Neurons in the Anterior Cingulate Cortex, in
The Journal of Neuroscience, 40(46), 8799-8815.
Gallero-Salas Yasir, Han Shuting, Sych Yaroslav, Voigt Fabian F., Laurenczy Balazs, Gilad Ariel, Helmchen Fritjof (2020), Sensory and Behavioral Components of Neocortical Signal Flow in Discrimination Tasks with Short-Term Memory, in
Neuron, 109, 1-14.
Banerjee Abhishek, Parente Giuseppe, Teutsch Jasper, Lewis Christopher, Voigt Fabian F., Helmchen Fritjof (2020), Value-guided remapping of sensory cortex by lateral orbitofrontal cortex, in
Nature, 585(7824), 245-250.
Haessig Germain, Milde Moritz B., Aceituno Pau Vilimelis, Oubari Omar, Knight James C., van Schaik André, Benosman Ryad B., Indiveri Giacomo (2020), Event-Based Computation for Touch Localization Based on Precise Spike Timing, in
Frontiers in Neuroscience, 14(420), 1-19.
Brains learn from experience by adjusting their behavioral strategies to optimize a desired output, e.g., reward collection or danger avoidance. A key learning mechanism is to evaluate the mismatch between an internal model of the world and the actual interaction and to update the internal model according to this prediction error. Prediction errors can be used to correct bottom-up (recognition) and top-down (generative) signal flows by adjusting synaptic weights in the neural circuitry. In addition, a more global and delayed prediction error encoding a mismatch of desired and actual outcome value (‘reward prediction error’) can be propagated through the brain network and induce synaptic adaptations as well. Although prediction error handling and error propagation have been studied for decades, the principles of how neuronal networks in the brain update the internal model through synaptic plasticity remain poorly understood. Understanding the core mechanisms of error processing would not only provide fundamental insight into the amazing adaptive behaviors seen in the animal kingdom but also open new exciting avenues for innovative brain-inspired fast and self-learning neuromorphic computing systems. Here, we propose a new theory-driven approach validated by neuroscientific experiments to identify common error-coding principles in two animal models, the mouse and the fruit fly, combined with engineering approaches to build physical computing systems that embody the identified learning principles. The two animal models offer complementary physiological and genetic tool kits necessary to test theoretical hypotheses. Both will crucially profit from experimental setups that include real-time feedback modulation with neuromorphic electronic devices. Our central hypothesis is that generative and recognition errors are jointly processed in distinct dendritic compartments of individual neurons as well as in specific subtypes of inhibitory interneurons controlling these compartments. We will test this hypothesis by measuring neural activity during conditioned behaviors in the mouse barrel cortex and the fly mushroom body, and by specifically manipulating error-coding elements using optogenetic tools. With low-latency neuromorphic (‘spiking’) sensors and computing hardware we will establish a novel real-time closed-loop brain stimulation system enabling direct interactions with the error handling processes, either impeding or improving behavioral adaptations. Our interdisciplinary collaboration with complementary expertise will lay the ground for developing more evolved computational hypotheses regarding biological mechanisms underlying learning as well as novel neuromorphic hardware that guides the way to future self-learning computing devices.