Cell Assemblies (CAs)
- CAs are recurrent networks that represent concepts
- A network may have several (millions) CAs
- A CA consists of a group of neurons that can maintain activity
within the group
- Connection strengths (synapses) between neurons within a CA tend to
be higher than connections between CAs
- Neurons can participate in more than one CA
- An active CA represents a short term memory. A CA is active when
it can sustain activity (reverberate) without external input
- A CA (something that can be activated) is a long-term memory
- CAs are formed via Hebbian learning
- They're not a new concept
- Hebb proposed them in 1949
- An untrained network is repeatedly presented with instances of
items
- The net learns to categorise these instances as CAs
- A CA or concept is a bunch of features that tend to travel around
together
- There are not necessary and sufficient conditions for membership
- Instead the categories are based around family resemblance
like Roschian categories [Rosch]
- CAs are pseudo-stable states that evolve via Hebbian learning
from presentations of multiple instances
- An instance that the net has not seen can be categorised
because it is similar to those that the network has seen
