Energy Aware Computing
Outline:

Until now, energy efficiency has not been considered a key aspect of AI, rather many AI solutions used "brute force" requiring huge amount of computing power to achieve the desired performance. Studying efficient energy consumption in brain's intelligence helps us to design AI that is energy efficient. In our approach, a system that wastes energy is not considered suitable to produce intelligence. We develop superior AI in which energy constraint leads to intelligence.

In our working model, spiking neuron populations are combined with metabolic equations in a unique way.

Click for more information

One salient point is the coupling of processes at very different time scales: fast neuron spiking at milliseconds scale, metabolic processes at time scale of fractions of a second, and vascular effects which are several orders of magnitude slower at the scale of 10 s or longer. We compare and explore specific advantages of these approaches when studying phase transitions as hallmarks of intelligent functions.

The Capillary-Astrocyte-Neuron (CAN) model is a two-compartment model with a metabolic component combining the cerebral blood flow (CBF) contribution with the astrocytic glycogen store and mitochondrial production of ATP, and a second component being the spiking neuron component.

Click for more information

Each individual “CAN” is modeled by one capillary, one glia cell (astrocyte), and one neuron. The end result is that we have constructed a model of a patch of cortical neuropil which simulates the complex feedback flow of metabolic energy through that patch along with a description of how that feedback flow affects the spiking behavior of the individualneuron in relation to the input influences impinging on it from other units in the local pool.

Schematic Diagram of Single Capillary-Astrocyte-Neuron Unit
Schematic Representation of Single Capillary-Astrocyte-Neuron Unit




Model Properties:

The model that was developed is capable of displaying many interesting behaviors. For example it is possible to cause the neurons to spike within different frequency bands by modulating the parameters in a certain way.

Click for more information

For the purpose of this project we optimized the model to obtain those range of parameters such that the neurons have a high degree of amplitude synchronization and we have the ability to modulate the amount of power in the gamma band. The gamma band is considered special because the brain waves observed in the brain during higher cognitive functions are often in gamma band.

Synchrony ROI I (same range of beta and eps)
The metric plotted against the model parameters (β, ε) in figure above, denoted as χ, represents the amplitude synchrony among the excitatory neurons. The darker regions represent higher synchrony among the firing activity of the different neurons and the lighter regions represent lower synchrony among the firing activity of the neurons.
Power in Gamma Band ROI I (same range of beta and eps)
The metric plotted against the model parameters (β, ε) in figure above represents the amount of power in the gamma band. The darker regions represent higher power in the gamma band and the lighter regions represent lesser power in the gamma band.

GUI Illustrating the operation of the coupled Spiking NN and Metabolic Model with spike density dependent on the energy input and synchronization as the function of the metabolic feedback on the spiking model.

Team members:

Publications (Please refresh at the destination if document does not load): Click the links below for additional details (Please refresh at the destination if document does not load):

References