Adaptive Inhibition for Optimal Energy Consumption by Animals, Robots and Neurocomputers

In contrast to artificial systems, animals must forage for food. In biology, the availability of energy is typically both precarious and highly variable. Most importantly, the very structure of organisms is dependent on the continuous metabolism of nutrients into ATP, and its use in maintaining homeostasis. This means that energy is at the centre of all biological processes, including cognition. So far, in computational neuroscience and artificial intelligence, this issue has been overlooked. In simulations of cognitive processes, whether at the neural level, or the level of larger brain systems, the constant and ample supply of energy is implicitly assumed. However, studies from the biological sciences indicate that much of the brain’s processes are in place to maintain allostasis, both of the brain itself and of the organism as a whole. This also relates to the fact that different neural populations have different energy needs. Many artificial systems, including robots and laptop computers, have circuitry in place to measure energy consumption. However, this information is rarely used in controlling the details of cognitive processing to minimize energy consumption. In this work, we make use of some of this circuitry and explicitly connect it to the processing requirements of different cognitive subsystems and show first how a cognitive model can learn the relation between cognitive ‘effort’, the quality of the computations and energy consumption, and second how an adaptive inhibitory mechanism can learn to only use the amount of energy minimally needed for a particular task. We argue that energy conservation is an important goal of central inhibitory mechanisms, in addition to its role in attentional and behavioral selection.