Shunting Inhibition's and Dendrites' Vital Role in Neuromorphic Computing

 

It has long been known that shunting inhibition (as caused by GABAergic synapse onto the dendritic tree) is critical in a individual neurons computations. Specifically it has long been known that shunting inhibition performs a type of dirty division on excitatory input distal of the shunting inhibition. This is because shunting inhibition does not cause any deviation from the rest potential (or very small deviation). Shunting inhibition effectively makes the leak conductance stronger. As such any excitatory current becomes divided by the now larger leak conductance.

To see this more explicitly consider a GABAergic synapse with a reversal potential near leak $E_{GABA} \approx E_L$, and an excitatory input $I_E$. We would then get 
$$ \frac{dV}{dt} =I_{E}- g_L (V-E_L) -g_{GABA}(V-E_{GABA})$$

 
If we consider the steady state solution of this we can see how much the excitatory synaptic current depolarizes the cell. This is a measure of how strong that input is
$$ \frac{dV}{dt} =0=I_{E}- g_L (V-E_L) -g_{GABA}(V-E_{GABA})$$
$$  0=I_{E}- (g_L+g_{GABA}) (V-E_L)$$

$$ \frac{I_{E}}{ (g_L+g_{GABA})}= (V-E_L)$$

Notice how as $g_{GABA}$ goes up in effect of $I_E$ goes down, as if it is being divided by the GABA strength. This is especially more accurate if $g_L$ is close to zero. This gives a neuron's dendrite a great way to perform division on any excitatory input.

Fascinatingly, Dr.  Frances Chance, and Dr. Suma Cardwell, have shown this dendritic shunting inhibition is also important for neuromorphic computing. while shunting inhibition can be done in individual neurons, by separating shunting inhibitions into a simple linear dendrite, Dr. Chance and Dr. Cardwell can minimize the number of components used to implement shunting inhibition. 

Recall that each extra neuron added to a neuromorhpic chip requires a decent amount of electronic components to capture the nonlinear spiking mechanism. However, a dendrite can be implemented with a transistor(GABA synapse/ Leak conductance) and a resistor (axial connection between soma and dendrite). This saves on the component number, and thus reduces power. This is critical in producing a more computationally complex neuromorphic circuits, while keeping power usage to a minimum. This work can result in neuromorphic chips becoming the computer chip for low-power applications.

Author: Alexander White

Citation: Shunting Inhibition as a Neural-Inspired Mechanism for Multiplication in Neuromorphic Architectures (2023). S.F Chance, S. G. Cardwell, NICE 2023, April 11–14, 2023, San Antonio, TX, USA.