E-I Net FAQ
What is E-I Net?
E-I Net is a neural circuit model composed of two populations of spiking
neurons: excitatory and inhibitory. Many models of neural networks
imagine that “nodes” can send either positive or negative values to other
nodes via “connection weights” that can also be positive
or negative. However actual neurons in the brain don’t work this way.
In actual neural circuits in the brain, nearly all the neurons send
one of two signal types: excitatory or inhibitory.
The excitatory neurons send only positive signals whereas the inhibitory
send only negative signals. Furthermore, the connection weight multipliers
at the synapses can never be negative. E-I Net models this
type of circuit — a closer approximation to how the brain really works.
Why are neural circuits important?
The brain is often described as being composed of networks of neurons,
however there is actually more structure to the brain than that. Neurons
of specific types are wired together in stereotypical circuit patterns
that are repeated millions or billions of times in a parallel
overlapping fashion. For many of these circuits, the wiring diagram
is known whereas how the circuit works is not. Modeling these wiring
patterns as repeating circuits rather than unstructured networks makes
it possible to understand the relationship between the wiring pattern
structures and the brain’s information processing capacities.
Why spiking neural circuits?
Neurons in the brain communicate by emitting spikes — momentary electrical
impulses lasting around 1 millisecond. Neurons on average emit around
10 spikes per second in irregular patterns that sound like a Geiger counter.
These spikes are the basic language of the brain. Many network models
assume “neurons” send “graded” or “real number” (floating point) values,
however we know for a fact that this is not what actual neurons do.
Spiking neural circuits model what neurons actually do and how they
process information on the millisecond time scale.
What does E-I Net do that is interesting?
E-I Net is able to detect patterns and learn representational schemes
on its own without being told what the right answer is (unsupervised learning).
It does this by discovering the statistically independent
components of the input signal and forming a “sparse code.” A sparse code
is a type of representation in which most of the neurons do nothing
most of the time. Not only does sparse coding describe how the brain
appears to represent information, but it also has some interesting
properties that may explain how the brain is able to make flexible
inferences about complex information patterns. Here are a couple web
pages on sparse coding:
In addition, the spike patterns produced by E-I Net have superficial
similarities to the irregular rhythmic patterns found in the brain.
Here are two representative spike rasters produced by E-I Net:
Back to the E-I Net page