In this paper, Eugene M. Izhikevich modeled a small network consisting of 1000 spiking neurons where 80% are excitory and 20% are inhibitory, with 0.1 probability of connection. The connection weights are subjected to STDP and conduction delays. The network shows a large number of polychronous groups (PGs). And the number can be bigger than the number of neurons, even bigger than the number of synapses!
The neural polychronous groups are used as the basis for representing concepts in Conceptual Neural Network, namely polychronous group(s) are associated with a concept. So we need to take a look at the amazing properties of PGs.
Izhikevich described the properties of PGs and the network in this paper by observing the simulated model.
1. PGs are formulated due to the existance of STDP, which potentiates or depresses some weights.
2. Since STDP is always “on”, PGs constantly appear and disappear, while some of them can exist and survive within the entire duration of simulation.
3. It is very common that PGs share neurons. These shared neurons could be activated in different time patterns, so that they won’t be ambiguous.
4. PGs are often partly activated. The beginning part is very precise while the tail sometimes doesn’t preserve the precision very well.
(very close to CNN)5. Different PGs can represent different input patterns. Note that it is also quite common that meta-PG groups encode an input pattern, while not a single group.
6. Different PGs can compete with each other, which is of neural Darwinsim. This is also a source of selective attention.
7. The network can exhibit spontaneous activation, which can be similar to external input. As the paper goes, “One can say that the network thinks of the stimulus – that is, it pays attention to memory of the stimulus”.
In most spiking neural network simulations, the delay between two connected neurons are always set to a constant due to, maybe, mathematical simplicity. It is known that the delay is also plastic and can be modulated by brain. I plan to make use of this tunable parameter, along with STDP on axonal efficacy (weight), to achieve better performance for CNN.